James W.

Kolari and Seppo Pynnönen

Investment Valuation and Asset Pricing

Models and Methods
James W. Kolari
Mays Business School, Texas A&M University, College Station, TX, USA

Seppo Pynnönen
Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

ISBN 978-3-031-16783-6 e-ISBN 978-3-031-16784-3

https://doi.org/10.1007/978-3-031-16784-3

© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2023

This work is subject to copyright. All rights are solely and exclusively
licensed by the Publisher, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any
other physical way, and transmission or information storage and retrieval,
electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks, service

marks, etc. in this publication does not imply, even in the absence of a
specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.

The publisher, the authors, and the editors are safe to assume that the advice
and information in this book are believed to be true and accurate at the date
of publication. Neither the publisher nor the authors or the editors give a
warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The
publisher remains neutral with regard to jurisdictional claims in published
maps and institutional affiliations.
This Palgrave Macmillan imprint is published by the registered company
Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham,
Switzerland
For my wife Karie and family Wes and Mary.
—James W. Kolari
For my wife Marja Leena and family Lauri, Erno,
and Laura and Nicklas and grandchildren
Sebastian, Benjamin, and Vanessa.
—Seppo Pynnönen
with admiration and love.
Preface
A major change in the field of finance was marked by the advent of the
Capital Asset Pricing Model (CAPM). Numerous authors are credited with
this famous model of how asset prices are determined, including Treynor
(1961, 1962), Sharpe (1964), Lintner (1965), Mossin (1966), and Black
(1972). According to Black (1981) and recounted by French (2003), Jack
Treynor is now recognized to have developed the first version of the CAPM
as early as 1958 during a summer vacation to Colorado. As a former
graduate of Harvard University, he shared his writings with John Lintner at
Harvard in 1960. Never formally publishing his work, in the early 1960s he
presented the CAPM as a graduate student at the Massachusetts Institute of
Technology (MIT) to finance faculty, including Robert Merton.1 A close
friend and colleague was Fischer Black, with whom numerous discussions
of the model likely led to Black's later work on the related zero-beta
CAPM. Additionally, Jan Mossin was from Norway but attended Ph.D.
studies at Carnegie Mellon University in the United States. Together, these
students and professors worked contemporaeously on the fundamental
relation between risk and return (or prices) of assets.
An important foundation underlying the CAPM is the mean-variance
analyses of Markowitz (1952, 1959) and Tobin (1958). Harry Markowitz
received his Ph.D. in Economics from the University of Chicago on
portfolio theory. His work led to the fundamental concept of diversification
in portfolio investments. By combining assets in a portfolio with returns
that are less than perfectly correlated, investors can reduce risk as measured
by the variance of returns. In theory, by varying expected returns, an infinite
number of portfolios comprised of assets can be constructed that have
minumum risk at different return levels. These portfolios trace out an
efficient frontier in expected return/variance space. Relevant to the CAPM,
total risk (or variance of returns) can be divided into diversifiable and
nondiversifiable risks. It is the nondiversifiable portion of risk, known as
systematic risk, that is the focus of the CAPM. Markowitz worked for some
time at the Cowles Foundation of Yale University, where he met James
Tobin, a Harvard trained economist and Yale professor. Tobin’s insightful
separation principle posited that risk-averse investors seek: (1) efficient
portfolios and (2) some optimum fraction of an efficient portfolio and a
riskless asset as determined by their risk aversion. Importantly, these two
investment decisions are separate from one another.
It is interesting that the lives of these early pioneers in asset pricing
crossed paths with one another during the development of the CAPM. Their
interaction was instrumental in the evolution of breakthroughs that
transformed our thinking about risk and return in finance. Of course, it
helps to be smart too! In this regard, Nobel Prizes in Economics were
awarded to Tobin (1981), Markowitz (1990), and Sharpe (1990). Other
contributors did not receive this honor due to death or nonpublication but
are no less appreciated for their intellectual achievements. Certainly all of
these notable scholars played important roles in shaping our understanding
of modern finance. This book pays tribute to their accomplishments by
retracing their steps in financial history.
Investments Valuation and Asset Pricing: Models and Methods is
intended to fill a gap in undergraduate finance curriculums by providing an
asset pricing text that is accessible to undergraduate students. The course is
most suitable for senior finance students in the last year of their
undergraduate studies. Also, it can be used in a graduate finance class,
including Executive M.B.A. and Executive Business Education. The book
has three unique features that set it apart from other finance textbooks.
1.
Original published studies by researchers on the foundations of asset
pricing are reviewed in chronological order over time. Retracing their
steps in financial history, each chapter stays close to their authentic
works, including quotations, examples, graphical exhibits, and
empirical results. We want undergraduate students to gain a firm grasp
of their achievements in the theory and practice of finance.
2.
Important statistical concepts and methods relevant to the field of
finance are covered. These statistical materials are crucial to learning
asset pricing, which often employs statistical tests to evaluate different
asset pricing models. Also, statistical skills are important tools for any
finance professional.
3.
Practical examples, questions, and problems are included in the text to
help students check their learning and better understand the
fundamentals of asset pricing.
 The roadmap for our journey through the emergence and development
of asset pricing is as follows.
Chapter 1 reviews the diversification principles of Markowitz (1952,
1959).
Chapter 2 introduces the basic assumptions of investor behavior and
financial markets that are foundational to asset pricing models.
Chapter 3 builds upon these principles by following Sharpe’s (1964)
derivation of the CAPM under equilibrium market conditions.
Chapter 4 covers the CAPM market model that is used for empirical
estimation using real world return data.
Chapter 5 reviews the first extension of the CAPM known as the zero-
beta CAPM by Black (1972).
Chapter 6 overviews a number of other forms of the CAPM developed by
researchers, including the intertemporal CAPM, international asset
pricing model (IAPM), consumption CAPM (CCAPM), production
CAPM (PCAPM), and conditional CAPM.
Chapter 7 reviews the famous arbitrage pricing model (APT) by Ross
(1976).
Chapter 8 introduces the innovative Fama and French (1993) three-factor
model.
Chapter 9 expands the discussion to other multifactor models that have
become increasingly popular over the past few decades.
Chapter 10 reviews theory and evidence with respect to a recently
proposed special case of the zero-beta CAPM dubbed the ZCAPM by
Kolari, Liu, and Huang (2021).
Chapter 11 applies asset pricing models to event studies that investigate
the effects of market news announcements on stock returns.

Acknowledgements The authors would like to express their sincere

gratitude to our Editor Tula Weis as well as Production Manager Redhu
Ruthroyoni at Springer. This book would not have been possible without
the support of Editor Weis, who worked with us in developing the proposal
and shaping the title and content of our book. A number of anonymous
referees of our proposal provided valuable comments that were useful in
writing the book. Of course, the entire team at Springer is very much
appreciated for all their work in making this dream possible.
References
Black, F. 1972. Capital market equilibrium with restricted borrowing.
Journal of Business 45: 444–454.
Black, F. 1981. An open letter to Jack Treynor. Financial Analysts
Journal July/August. Letters to the Editor, p. 14.
Fama, E.F., and K.R. French. 1993. The cross-section of expected
returns. Journal of Financial Economics 33: 3–56.
French, C. W. 2003. The Treynor capital asset pricing model. Journal of
Investment Management 1: 60–72.
Kolari, J.W., W. Liu, and J.Z. Huang. 2021. A New Model of Capital
Asset Prices: Theory and Evidence. Cham, Switzerland: Palgrave
Macmillan.
Korajczyk, R.A. 1999. Asset Pricing and Portfolio Performance, 15–22.
London, UK: Risk Books.
Lintner, J. 1965. The valuation of risk assets and the selection of risky
investments in stock portfolios and capital budgets. Review of Economics
and Statistics 47: 13–37.
Markowitz, H.M. 1952. Portfolio selection. Journal of Finance 7: 77–91.
Markowitz, H. M. 1959. Portfolio Selection: Efficient Diversification of
Investments. New York, NY: John Wiley & Sons.
Mossin, J. 1966. Equilibrium in a capital asset market. Econometrica 34:
768–783.
Ross, S. A. 1976. The arbitrage theory of capital asset pricing. Journal of
Economic Theory 13: 341–360.
Sharpe, W. F. 1964. Capital asset prices: A theory of market equilibrium
under conditions of risk. Journal of Finance 19: 425–442.
Tobin, J. 1958. Liquidity preference as behavior toward risk. Review of
Economic Studies 4: 65–86.
Treynor, J. L. 1961. Market value, time, and risk. Unpublished
manuscript.
Treynor, J. L. 1962. Toward a theory of market value of risky assets.
Unpublished manuscript.
James W. Kolari
Seppo Pynnönen
College Station, USA
Vaasa, Finland
Contents
1 Portfolio Theory and Practice
1.1 Ex Ante Valuation
1.2 Ex Post Returns and Risk
1.3 Summary
Questions
Problems
Appendix A: Optimal Weights for Two Securities
Appendix B: Optimal Weights for Many Assets
References
2 Capital Market Conditions
2.1 Perfect Capital Markets
2.2 Efficient Markets
2.3 Risk Aversion
2.4 Normally Distributed Returns
2.5 Summary
Questions
Problems
References
3 Capital Asset Pricing Model (CAPM)
3.1 The Present Value of Future Cash Flows
3.2 Market Equilibrium Assumptions
3.3 Investment Choice with a Riskless Asset
3.4 The CAPM
3.5 Implications of the CAPM
3.6 Summary
 Questions
Problems
References
4 The Market Model
4.1 Empirical Form of the CAPM
4.2 Early Tests of the Market Model
4.2.1 Black, Jensen, and Scholes Study
4.2.2 Fama and MacBeth Study
4.3 Estimating the Market Model
4.4 A Primer on Regression Analysis
4.4.1 On the Interpretation of Regression Coefficients
4.4.2 Justification for OLS
4.5 OLS Estimation of the Market Model
4.5.1 Goodness-of-Fit
4.6 Summary
Questions
Problems
Appendix A: Statistical Inference
References
5 The Zero-Beta CAPM
5.1 Zero-Beta CAPM
5.1.1 Assumptions
5.1.2 Theoretical Model
5.2 Empirical Tests
5.3 Connection to Later Research
5.4 GRS Test for Portfolio Efficiency
 5.5 Summary
Questions
Problems
Appendix A: Mathematical Derivation of the Zero-Beta CAPM
Appendix B: R Snippet for the GRS Testing Example in Section 5.4
References
6 Alternative CAPM Specifications
6.1 Intertemporal CAPM
6.2 International Asset Pricing Model (IAPM)
6.2.1 World Market Factor
6.2.2 Exchange Rate Risk
6.3 Consumption CAPM
6.4 Production CAPM
6.5 Conditional CAPM
6.6 Summary
Questions
Problems
References
7 The Arbitrage Pricing Theory Model
7.1 APT Model
7.1.1 Assumptions
7.1.2 Theoretical Model
7.2 Empirical Testing
7.3 Summary
Questions
Problems
 References
8 Multifactor Models
8.1 What Explains Stock Returns?
8.2 Fama and French Three-Factor Model
8.3 Three-Factor Controversy
8.4 Summary
Questions
Problems
References
9 Anomalies and Multifactor Models
9.1 Carhart Four-Factor Model
9.2 Fama and French Four- and Five-Factor Models
9.3 Hou, Xue, and Zhang Four-Factor Model
9.4 Stambaugh and Yuan Four-Factor Model
9.5 Fama and French Six-Factor Model
9.6 Fama and French Cross-Section Factors
9.7 Lettau and Pelger Latent Five-Factor Model
9.8 Rise of the Machines
9.9 Summary
Questions
Problems
References
10 A Special Case of the Zero-Beta CAPM: The ZCAPM
10.1 Theoretical ZCAPM
10.2 Empirical ZCAPM
10.3 Empirical Evidence
 10.3.1 Risk and Return Relations
10.3.2 Predicted and Actual Return Relations
10.3.3 Fama and MacBeth Cross-Sectional Tests
10.4 Summary
Questions
Problems
References
11 Event Studies
11.1 Short-Run Event Studies
11.1.1 Basic Steps in an Event Study
11.1.2 Abnormal Returns
11.1.3 Parametric Testing
11.1.4 Examples
11.2 Long-Run Event Studies
11.2.1 Buy-and-Hold Abnormal Return Approach
11.2.2 Calendar Time Abnormal Return Approach
11.2.3 Example
11.3 Summary
Questions
Problems
Appendix A: Nonparametric Testing
References
Index
List of Figures
Fig. 1.1 Correlation of security returns and effects on portfolio returns

Fig. 1.2 Combinations of two securities to form a portfolio

Fig. 1.3 The minimum variance boundary of portfolios based on data for
securities 1 and 2 in Table 1.3

Fig. 1.4 The minimum variance boundary of portfolios and opportunity set
of all assets

Fig. 2.1 Utility analysis of choices involving risk

Fig. 2.2 Expected returns and utility for multiple risky investment choices

Fig. 2.3 Mean-variance indifference curve of a risk-averse investor

Fig. 2.4 Empirical return distributions for Microsoft, U.S. S &P 500 index,
and German Dax index using daily, weekly, and monthly returns

Fig. 3.1 Equilibrium relationship between expected rates of return and risk

Fig. 3.2 Combining the efficient frontier and indifference curves

Fig. 3.3 Locating market portfolio as a tangent point on the ray from
riskless rate

Fig. 3.4 Pricing capital assets using systematic or beta risk

Fig. 3.5 Expected rates of return and beta risk

Fig. 4.1 The estimated market model for Microsoft (MSFT) stock returns
with fitted regression line

Fig. 5.1 A graphical depiction of Black’s zero-beta CAPM

Fig. 5.2 Industry portfolios and S &P 500 index returns with ex post
efficient frontier and the market line

Fig. 6.1 A three-dimensional depiction of an empirical ICAPM with two

systematic beta risk factors: (1) market factor beta , and (2) state
variable beta

Fig. 6.2 Alternative CAPM specifications based on the intertemporal

CAPM (ICAPM)

Fig. 6.3 Currency risk in world currency markets affects the stock returns in
different countries through their exchange rate risk exposure or sensitivity
to national currency movements
Fig. 6.4 Consumers buy and sell financial assets over time to smooth
consumption in the consumption CAPM (CCAPM)

Fig. 6.5 Close relation between stock return and capital investment
forecasts in the production CAPM (PCAPM)

Fig. 6.6 Beta estimates for the market factor tend to vary randomly over
time consistent with the unconditional CAPM but contrary to the
conditional CAPM

Fig. 7.1 Arbitraging riskless profits by buying and selling mispriced assets
(Source Adapted from Wikipedia
[https://en.wikipedia.org/wiki/Arbitrage_pricing_theory])

Fig. 8.1 Stocks are sorted into a 2 3 matrix by size and value (BM). The
size factor is the average returns of the three small (blue) portfolios minus
the three (red) big portfolios, and the value factor is average returns of the
two value (green) portfolios minus the two growth (brown) portfolios

Fig. 8.2 The Fama and French three-factor model with excess returns a
function of market, size, and value factors takes into account four
dimensions in return/risk space

Fig. 9.1 Momentum factor returns are long/short zero-investment portfolio

returns

Fig. 9.2 Multifactor models extensions of the original Fama and French
three-factor model using discretionary methods
Fig. 9.3 Machine learning methods based on artificial intelligence (AI)
models rather than human judgment

Fig. 10.1 Locating orthogonal portfolios and on the mean-variance

parabola (Source Kolari et al. 2021, p. 59)

Fig. 10.2 The ZCAPM with beta risk related to average excess market
returns and zeta risk associated with positive and negative sensitivity to
return dispersion (Source Kolari et al. 2021, p. 72)

Fig. 10.3 Two-sided, opposite effects of return dispersion on the expected

returns of two assets (Source Kolari et al. 2021, p. 94)

Fig. 10.4 This graph shows the average beta risk, zeta risk, and one-month-
ahead equal-weighted returns for 25 beta-zeta sorted portfolios. These long-
only portfolios approximate the shape on an investment parabola. In each
one-year estimation window, the empirical ZCAPM is estimated using daily
returns to obtain the beta and zeta risk parameters. The analysis period is
January 1965–December 2018 (Source Kolari et al. 2021, p. 272)

Fig. 10.5 Out-of-sample cross-sectional ZCAPM relationship between

average one-month-ahead realized excess returns in percent (Y-axis) and
average beta risk in the previous 12-month estimation period (X-axis).
Results are shown for 25 size-BM portfolios used in the Fama and French
studies. Portfolios are sorted into zeta risk $$ Z^*_{i,a} $$ quintiles and
then beta risk quintiles within each $$ Z^*_{i,a} $$ quintile. The
analysis period is January 1965–December 2018 (Source Kolari et al. 2021,
p. 141)
Fig. 10.6 Out-of-sample cross-sectional ZCAPM relationship between
average one-month-ahead realized excess returns in percent (Y-axis) and
average zeta risk $$ Z^*_{i,a} $$ in the previous 12-month estimation
period (X-axis). Results are shown for 25 size-BM portfolios used in the
Fama and French studies. Portfolios are sorted into beta risk
$$ \beta _{i,a} $$ quintiles and then zeta risk $$ Z^*_{i,a} $$ quintiles
within each $$ \beta _{i,a} $$ quintile. The analysis period is January
1965–December 2018 (Source Kolari et al. 2021, p. 140)

Fig. 10.7 Out-of-sample cross-sectional relationship between average one-

month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25
size-BM sorted portfolios: Fama and French three-factor model in Panel A
and empirical ZCAPM in Panel B. The analysis period is January 1965 to
December 2018 (Source Kolari et al. 2021, p. 146)

Fig. 10.8 Out-of-sample cross-sectional relationship between average one-

month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25
size-BM sorted plus 47 industry portfolios: Fama and French Fama and
French three-factor model in Panel A and empirical ZCAPM in Panel B.
The analysis period is January 1965–December 2018 (Source Kolari et al.
2021, p. 147)

Fig. 10.9 Out-of-sample cross-sectional relationship between average one-

month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25
profit-investment sorted portfolios: Fama and French six-factor model in
Panel A and empirical ZCAPM in Panel B. The analysis period is January
1965 to December 2018 (Source Kolari et al. 2021, p. 150)
Fig. 10.10 Out-of-sample cross-sectional relationship between average one-
month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25
profit-investment sorted plus 47 industry portfolios: Fama and French Fama
and French six-factor model in Panel A and empirical ZCAPM in Panel B.
The analysis period is January 1965–December 2018 (Source Kolari et al.
2021, p. 151)

Fig. 11.1 Cumulative abnormal returns for selected industries in response to

2008 financial crisis news events

Fig. 11.2 Cumulative abnormal returns for selected industries in response to

2020 COVID-19 crisis news events
List of Tables
Table 1.1 Expected rate of return calculation (in percent terms)

Table 1.2 Actual or realized (ex post) return and risk over time (in percent
terms)

Table 1.3 Portfolio return and risk effects (in percent terms)

Table 2.1 Sample statistics for S &P 500 index, DAX index, and Microsoft
Corporation: daily, weekly, and monthly returns (in percent)

Table 4.1 Monthly returns for Microsoft (MSFT), S &P 500 index, and
Treasury bills from January 2018 to December 2020

Table 8.1 Average monthly returns for U.S. stock portfolios sorted by size
and book-to-market equity (BM) from July 1963 to December 1990 in the
Fama and French (1992) study

Table 8.2 Goodness-of-fit $$ R^2 $$ estimates for 25 U.S. stock

portfolios sorted on size and book-to-market equity (BM) from July 1963 to
December 1991 in the Fama and French (1993) study

Table 8.3 Intercept or $$ \alpha $$ estimates for 25 U.S. stock portfolios

sorted on size and book-to-market equity (BM) from July 1963 to
December 1991 in the Fama and French (1993) study
Table 9.1 The relation between mutual fund performance and their
operating characteristics in the Carhart (1997) study

Table 9.2 Intercept or $$ \alpha $$ estimates for selected anomaly

portfolios from January 1967 to December 2013 in the Stambaugh and
Yuan (2017) study

Table 9.3 Mispricing error of different models in the Lettau and Pelger
(2020) study

Table 10.1 Out-of-sample Fama–MacBeth cross-sectional tests for ZCAPM

regression factor loadings compared to other asset pricing models in the
period January 1965–December 2018: 12-month rolling windows

Table 11.1 Short-run event study results for 50 SEO events of U.S. firms
from 1985 to 2015

Table 11.2 Abnormal returns for four industries using the Fama and French
five-factor model: Selected news announcement dates during the 2008
financial crisis and 2020 COVID-19 crisis

Table 11.3 Long-run event study results for 200 SEO events of U.S. firms
from 1985 to 2015
Footnotes
1 See Chapter 2 of Korajczyk (1999) for a reproduction of Treynor (1962).
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_1

1. Portfolio Theory and Practice

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX, USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa, Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1007/978-3-031-16784-3_1

This chapter overviews the portfolio selection process developed by Nobel Laureate Harry Markowitz (1952,
1959). Basic concepts are introduced that are important in the field of asset pricing. These concepts are
fundamental to understanding the relation between return and risk, which is the central theme of finance.
Markowitz showed how to reduce risk by diversification in assets with less than perfect correlation. Many experts
agree that diversification is the only “free lunch” available to investors in the market for real and financial assets.
Diversification leads to the mean-variance parabola that describes the lowest risk portfolios for any given return.
These major advances in modern portfolio theory have laid the foundation for general equilibrium asset pricing
models in forthcoming chapters. Moreover, they have revolutionized investment management practices around the
world.

1.1 Ex Ante Valuation

Markowitz viewed portfolio selection as a two-step process. First, investors assess the potential future performance
of securities. Because it is expected or anticipated in the future, we refer to this analysis as ex ante. Second, they
use this information to choose a portfolio of securities. Focusing on the second step, what rule should an investor
use to guide their portfolio choices? A simple rule is to maximize the discounted value of future investment returns
using the present value formula. As an example, assuming a three-year investment, the present value for a risky
security can be written as:
$$\begin{aligned} V =\frac{\$ 10}{(1+0.10)} + \frac{\$ 12}{(1+0.10)^2} + \frac{\$ 8}{(1+0.10)^3}
(1.1)
\approx \$25, \end{aligned}$$
where risky cash flows in each future year are shown in the numerator, and the discount rate is 10%. The discount
rate is the required rate of return of investors. We can re-write the formula in general terms as:
$$\begin{aligned} V = \frac{E(CF_1)}{[1+E(R)]} + \frac{E(CF_2)}{[1+E(R)]^2} + \frac{E(CF_3)}
(1.2)
{[1+E(R)]^3}, \end{aligned}$$
where $$E(CF_t)$$ = the expected cash flow at time t, and E(R) = the expected rate of return or discount rate in
the three-year investment period. The expected value operator is used here to make clear that future cash flows are
not known with certainty and therefore are risky. Given expected cash flows, discount rate E(R) sets their present
value equal to the current price of the investment. Investors form expectations about uncertain cash flows based on
available information gathered about the firm, industry, and general economy.
The discount rate takes into account the risk of uncertain cash flows. Assuming that investors are risk averse
(or do not like risk), higher risk implies a higher discount rate and, in turn, lower present value or asset price. That
is, investors demand a higher expected return as compensation for the risk that they bear. If there is no risk of
future cash flows, the discount rate is the riskless rate of interest. U.S. government bond rates are often used to
proxy the riskless rate. Hence, the expected risky discount rate has two components:
$$\begin{aligned} E(R) = R_f + RP, \end{aligned}$$ (1.3)
where $$R_f$$ is the riskless rate, and RP is the risk premium. The latter risk premium increases as: (1) the
risk of the security increases, and (2) investors become more risk averse.
Let’s take a closer look at how investors could evaluate expected rates of return. Since the future is uncertain,
we do not know exactly what the cash flows will be over time. However, investors can estimate probabilities of
different possible outcomes or states. In Table 1.1 we provide an example of how to compute the expected
return for security j. Assume that there are three possible future states: (1) pessimistic, (2) average, and (3)
optimistic. The states $$s = 1$$, 2, 3 are assigned probabilities $$p_s = 0.25$$, 0.50, 0.25 and possible
returns $$R_{sj} = 5\%$$, $$10\%$$, $$15\%$$, respectively. Let’s assume that we hold the
security for one year. The expected annual return is computed as:
$$\begin{aligned} E(R_j) = \sum _{s= 1}^S p_s R_{sj}, \end{aligned}$$ (1.4)
which equals 10% in Table 1.1.
Table 1.1 Expected rate of return calculation (in percent terms)

State $$p_s = \text $$R_{sj $$p_sR_{sj}$$ $$R_{sj}- $$[R_{sj} - $$p_s[R_{sj} -

{probability}$$ }$$ E(R_{sj})$$ E(R_{sj})]^2$$ E(R_{sj})]^2$$
1 0.25 5 1.25 −5 25 6.25
2 0.50 10 5.00 0 0 0.00
3 0.25 15 3.75 5 25 6.25
1.00 $$E(R_j) = \text $$\sigma ^2(R_j) = \text
{10}$$ {12.50}$$
$$\sigma (R_j) = \text
{3.5355}$$

Should investors seek securities that maximize expected returns? On its surface, the answer to this simple
question would seem to be “yes.” However, Markowitz (1952, p. 77) argued that, “The hypothesis (or maxim) that
the investor does (or should) maximize discounted return must be rejected.” In his words, any rule of investment
behavior that ignored “... the superiority of diversification must be rejected.” A better rule is to maximize expected
returns per unit of risk. Risk can be conveniently measured as the variance of returns.
Returning to Table 1.1, the variance of returns across states for security j is equal to 12.50. Variance is a
measure of the dispersion, volatility, or variability of returns across the states. The formula for the variance of
returns is:
$$\begin{aligned} \sigma ^2(R_j) = \sum _{s=1}^S p_s [R_{sj}-E(R_j)]^2. \end{aligned}$$ (1.5)
The variance of returns is not in percentage terms like expected returns. However, the standard deviation of
returns is an equivalent measure of total risk that has units of measurement in percentage terms like expected
returns. We can easily compute the standard deviation of returns as:
$$\begin{aligned} \sigma (R_j)= \sqrt{\sigma ^2(R_j)}. \end{aligned}$$ (1.6)
Using the data in Table 1.1, we get approximately $$\sigma (R_j)=\sqrt{12.50}\approx 3.54\%$$. Given
a one standard deviation change in returns due to volatility, returns can vary around the expected return from a low
of $$6.46\%\ (= 10\% - 3.54\%)$$ to a high of $$13.54\%\ (= 10\% + 3.54\%)$$. A two-standard
deviation change implies a range from 2.92% to 17.08%. From statistics, it is known that two-standard deviations
encompass all possible future return outcomes with a 95% probability. This wide range of values suggests that
security j is fairly risky, which explains the relatively high expected return of 10%.
Based on these return and risk concepts, Markowitz established straightforward methods for evaluating
portfolio effects. The expected return of a portfolio is defined as:
$$\begin{aligned} E(R_P)= w_1 E(R_1) + w_2 E(R_2), \end{aligned}$$ (1.7)
where $$w_1$$ is the proportion invested in security 1, and $$w_2$$ is the proportion invested in
security 2. Note that the sum $$w_1 + w_2 =1$$, i.e., $$w_2 = (1-w_1)$$. What is the variance of
portfolio returns? Here is where Markowitz made an important contribution to our thinking about risk. The
variance formula from basic statistical methods is:
$$\begin{aligned} \sigma ^2(R_P) = w^2_1 \sigma ^2(R_1) + 2 w_1 w_2 Cov (R_1,R_2) + w^2_2 \sigma
(1.8)
^2(R_2), \end{aligned}$$
where $$Cov (R_1,R_2)$$ = the covariance of returns between security 1 and security 2 over time.
Covariance can be expressed in terms of deviations from expected returns for each security:
$$\begin{aligned} Cov (R_1,R_2) = E\{[R_1 - E(R_1)] [R_2 - E(R_2)]\}. \end{aligned}$$ (1.9)
A useful statistical definition of return covariance is:
$$\begin{aligned} Cov (R_1,R_2) = \rho _{12} \sigma (R_1) \sigma (R_2), \end{aligned}$$ (1.10)
where $$\rho _{12}$$ = the correlation of returns between security 1 and security 2 over time. For two
securities, denoting the covariance term $$Cov (R_j,R_k)$$ in the simpler notation $$\sigma _{jk}$$,
we can re-write Eq. (1.8) in matrix form as:
$$\begin{aligned} \sigma _P^2 =(w_1, w_2) \begin{pmatrix} \sigma _{11} &{} \sigma _{12}\\
\sigma _{21} &{} \sigma _{22} \end{pmatrix} \begin{pmatrix} w_1\\ w_2 \end{pmatrix} = w^2_1 (1.11)
\sigma ^2_1 + w_1 w_2 (\sigma _{12} + \sigma _{21}) + w^2_2 \sigma ^2_2, \end{aligned}$$
where $$w_1 + w_2 = 1$$, $$\sigma ^2(R_1) = \sigma _{11} = \sigma ^2_1$$, and
$$\sigma ^2(R_2)= \sigma _{22} = \sigma ^2_2$$. Notice that the covariance terms $$\sigma _{12}$$
and $$\sigma _{21}$$ are equal to one another and, therefore, their sum equals $$2 \sigma _{12}$$
such that:
$$\begin{aligned} \sigma _P^2 = w^2_1 \sigma ^2_1 + 2 w_1 w_2 \sigma _{12} + w^2_2 \sigma ^2_2,
(1.12)
\end{aligned}$$
which is the same as Eq. (1.8).

1.2 Ex Post Returns and Risk

Table 1.2 Actual or realized (ex post) return and risk over time (in percent terms)

Years $$R_1$$ $$R_2$$ $$R_1 - $$[R_1 - E(R_1)]^2$$ $$R_2 - $$[R_2 - E(R_2)]^2$$

E(R_1)$$ E(R_2)$$
1 7 7 −5.2 27.04 0.4 0.16
2 13 8 0.8 0.64 1.4 1.96
3 8 5 −4.2 17.64 −1.6 2.56
4 16 3 3.8 14.44 −3.6 12.96
5 $$\underline{17 $$\underline{10 4.8 23.04 3.4 11.56
}$$ }$$
$$\bar{R}_1$$ = $$\bar{R}_2$$ = $$\sigma ^2(R_1)$$ = $$\sigma ^2(R_2)$$ =
12.2 6.6 82.8/4 = 20.70 29.20/4 = 7.30
$$\sigma (R_1)$$ = 4.55 $$\sigma (R_2)$$ = 2.70

Instead of assessing future expected returns, which requires dealing with uncertainty about the future, let’s
consider how to measure expected returns and risk from ex post or historical data. We can measure past returns and
risk over time. Suppose securities 1 and 2 have actual or realized returns in previous years 1 to 5 after investment
as shown in Table 1.2. To compute the sample mean or average return for security j over $$t = 1, \ldots , T$$
periods, we can use the following formula:
$$\begin{aligned} \bar{R}_j= \frac{\sum _{t=1}^T R_{jt}}{T}. \end{aligned}$$ (1.13)
The sample variance of returns over time t related to risk is:
$$\begin{aligned} \sigma ^2(R_j)= \frac{\sum _{t=1}^T (R_{jt} - \bar{R}_j)^2}{T-1}. \end{aligned}$$ (1.14)
 As shown in Table 1.2, the returns for securities 1 and 2 over the past five years are known. From this data we
can compute their average returns over the five years—namely, $$\bar{R}_1 = 12.2$$% and
$$\bar{R}_2= 6.6$$%, respectively. The variance of their returns can be computed, which yields
$$\sigma ^2(R_1) = 20.70$$ with $$\sigma (R_1) = 4.55$$% and $$\sigma ^2(R_2) = 7.30$$
with $$\sigma (R_2) = 2.70$$%.
This information can be utilized to see how the average portfolio return $$\bar{R}_P$$ and variance of
portfolio returns $$\sigma ^2(R_P)$$ changes as the proportion investment weights $$w_1$$ and
$$w_2$$ are changed. Table 1.3 shows the portfolio results. A naive investor that sought to maximize
average portfolio returns would choose weights $$w_1 = 1$$ and $$w_1 = 0$$ to earn an average
return ( $$\bar{R}_P$$) equal to 12.20%; hence, such investor would put all their funds in security 1.
However, according to Markowitz, we need to evaluate how the security weights affect both
$$\bar{R}_P$$ and $$\sigma (R_P)$$. Table 1.3 shows that, as $$w_1$$ increases from 0.1 to
0.9 and $$w_2$$ decreases from 0.9 to 0.1, the portfolio variance increases. If we take the ratio of average
returns/standard deviation of returns, we see that this ratio reaches a maximum at 8.28/2.25 = 3.68 using weights
$$w_1 = 0.3$$ and $$w_2 = 0.7$$. Investing only in security 1, such that $$w_1 = 1.0$$ and
$$w_2 = 0.0$$, this ratio equals 12.20/4.55 = 2.68, which is not as high as 3.68 for the two-security portfolio
with $$w_1 = 0.30$$ and $$w_2 = 0.70$$. It is clear that, per unit standard deviation of returns (or
risk), the investor is better off holding a portfolio containing both securities 1 and 2 with weights 0.30 and 0.70,
respectively. This example demonstrates that investment returns should only be evaluated in the context of
portfolio returns per unit risk. Investors should seek to maximize portfolio returns per unit risk as measured by the
standard deviation or variance of returns.
Table 1.3 Portfolio return and risk effects (in percent terms)

$$w_ $$w_ $$\bar{R}_ $$\bar{R}_ $$\bar{R}_ $$\sig $$\sig $$\r $$Cov $$\sigm $$\sig $$\bar
1$$ 2$$ 1$$ 2$$ P$$ ma ma ho (R_1,R_2)$$ a ma ma (R_P
(R_1)$$ (R_2)$$ _{12}$$ ^2(R_P)$$ (R_P)$$
0.0 1.0 12.2 6.6 6.60 4.55 2.70 0.1708 0.0000 7.29 2.70 2.44
0.1 0.9 12.2 6.6 7.16 4.55 2.70 0.1708 2.0983 4.40 2.10 3.41
0.3 0.7 12.2 6.6 8.28 4.55 2.70 0.1708 2.0983 5.05 2.25 3.68
0.5 0.5 12.2 6.6 9.40 4.55 2.70 0.1708 2.0983 7.40 2.72 3.46
0.7 0.3 12.2 6.6 10.52 4.55 2.70 0.1708 2.0983 11.45 3.38 3.11
0.9 0.1 12.2 6.6 11.64 4.55 2.70 0.1708 2.0983 17.19 4.15 2.80
1.0 0.0 12.2 6.6 12.20 4.55 2.70 0.4549 0.0000 20.70 4.55 2.68

In the above portfolio analyses, portfolio risk can be dramatically affected by the correlation between security
1 returns and security 2 returns, which can be measured by the correlation coefficient $$\rho _{12}$$. If
$$\rho _{12} < 1$$ , Markowitz argued that the total risk of the portfolio as measured by
$$\sigma ^2(R_P)$$ can be reduced. This reduction in risk is due to a diversification benefit arising from
less than perfect positive correlation at $$\rho _{12} = 1$$. The old adage that to diversify you should not
“... put all your eggs in one basket” needed amendment. According to Markowitz’s insight, to successfully
diversify, you need to “... put your money in multiple securities that are not perfectly correlated with one another.”
Using the simpler notation in Eq. (1.12), if securities 1 and 2 are perfectly positively correlated such that
$$\sigma _P^2 = w^2_1 \sigma ^2_1 +2w_1 w_2 \sigma _1 \sigma _2 (i.e.,
$$\rho _{12} = 1$$, then
+ w^2_2 \sigma ^2_2$$
$$\sigma _{12} = 1 \sigma _1 \sigma _2$$). If they are uncorrelated with $$\rho _{12} = 0$$ so that
$$\sigma _{12} = 0$$, then $$\sigma _P^2 = w^2_1 \sigma ^2_1 + w^2_2 \sigma ^2_2$$, as the term
$$2w_1 w_2 \sigma _{12}$$ = 0. Clearly, dropping this covariance term reduces the variance of portfolio
returns. Surprisingly, if securities 1 and 2 have perfectly negatively correlated returns with
$$\sigma _P^2 = w^2_1 \sigma ^2_1 - 2w_1 w_2 \sigma _1 \sigma _2
$$\rho _{12} = -1$$, such that
+ w^2_2 \sigma ^2_2$$
(i.e., $$\sigma _{12} = -1 \sigma _1 \sigma _2$$), portfolio variance can be substantially reduced.
Remarkably, assuming that the investment proportions $$w_1$$ and $$w_2$$ are chosen optimally,
perfect negative correlation of two securities can drive total risk down to zero! Appendix A shows how to compute
the optimal weights for two securities.

Fig. 1.1 Correlation of security returns and effects on portfolio returns

Figure 1.1 provides graphical depictions of how the correlation between the returns of securities 1 and 2 over
time affects portfolio returns. The top panel shows that, if the securities’ returns are perfectly positively correlated
with $$\rho _{12} = 1$$, the combined portfolio exhibits an identical pattern of returns over time. If
$$\rho _{12} = -1$$, portfolio returns could be flat line indicating zero total risk or volatility. In the more
likely case that $$0< \rho _{12} < 1$$, the portfolio return is less volatile than either security due to
diversification that results in risk reduction.
In the real world, it is most likely that $$0< \rho _{12} < 1$$. Indeed, it can be difficult to find
securities that have negative correlations with other securities over time. Given $$\rho _{12}$$, the investor
should seek to minimize the variance of portfolio returns by optimally choosing the investment proportions or
weights $$w_1$$ and $$w_2$$ to achieve this goal. These optimal weights can be solved to generate
the average return and variance of a minimum variance portfolio, or $$\bar{R}_P$$ and
$$\sigma ^2_P$$. In this regard, at each level of $$\bar{R}_P$$, a different minimum variance
portfolio is possible.

Fig. 1.2 Combinations of two securities to form a portfolio

Figure 1.2 illustrates the basic principles of Markowitz diversification. We assume that only two securities are
considered. Securities 1 and 2 have correlation somewhere in the reasonable range
$$0< \rho _{12} < 1$$. Average returns are on the Y-axis, and standard deviations of returns are on the
X-axis. The locations of securities 1 and 2 are shown in return/standard deviation space. The set of portfolios that
would exist if these securities were perfectly positively correlated ( $$\rho _{12} = 1$$) is shown by the
dashed line. Two other dashed lines coincide with different combinations of these securities if they were perfectly
negatively correlated. Portfolio X is the zero variance combination of these portfolios at some optimal weights.
Here we interchangeably use the words variance and standard deviation of returns, which measure total risk.
The curved bold line in Fig. 1.2 shows the infinite number of optimal combinations of securities 1 and 2 with
minimum variance based on varying the average return of the portfolio. Markowitz referred to this curve as the
mean-variance parabola. The range of average portfolio returns runs from all security 1 (i.e., $$w_1 = 1$$
so the $$\bar{R}_P= R_1$$) to all security 2 (i.e., $$w_2 = 1$$ so the $$\bar{R}_P = R_2$$).
Portfolio G is the global minimum variance portfolio that has the smallest variance of returns for all possible
combinations of these two securities. Portfolio G has average return $$R_G$$. Importantly, Markowitz
named the set of minimum variance portfolios on the upward sloping boundary of the bold curve the efficient
frontier. These portfolios are efficient in the sense that they have: (1) the lowest variance for any given average
portfolio return, and (2) the highest average portfolio return for any given variance of returns. He argued that
rational, risk-averse investors would only be interested in these efficient portfolios. Other portfolios on the
downward sloping boundary of the bold curve are inefficient minimum variance portfolios.

Fig. 1.3 The minimum variance boundary of portfolios based on data for securities 1 and 2 in Table 1.3

Figure 1.3 shows the efficient frontier based on security 1 and security 2 data for average portfolio returns and
standard deviations of portfolio returns in Table 1.3. Recalling this table, we varied the security weights
$$w_1$$ and $$w_2$$ from 0 to 1 for each security. The returns of securities 1 and 2 equal to 8.40%
and 9.60%, respectively, are represented by the endpoints of the bold frontier. Since their correlation coefficient
$$\rho _{12} = 0.3544 < 1$$, different combinations of securities 1 and 2 generate a Markowitzian
efficient frontier. Referring to the data in Table 1.3, the global minimum variance portfolio G has weights of
approximately $$w_1 \approx 0.1$$ and $$w_2 \approx 0.90$$. Portfolios on the minimum variance
boundary from G to security 1 comprise the efficient frontier. Investors would choose different minimum variance
portfolios based on their risk preferences. The most risk averse investors would logically choose G. As risk
aversion decreases, investors would choose portfolios with higher average returns on the efficient frontier.

Fig. 1.4 The minimum variance boundary of portfolios and opportunity set of all assets

Lastly, consider the more realistic and general case of many assets, not just securities but all assets. The
population of assets includes both financial and nonfinancial (real) assets such as stocks, bonds, real estate,
commodities, etc. As the number of assets under consideration increases, the efficient frontier will gradually move
left due to increasing diversification benefits. Figure 1.4 illustrates the efficient frontier for all assets in the
opportunity set at a single point in time.
It is interesting to consider what happens to the shape of the parabola over time. Some assets may experience
higher returns and others lower returns than in previous periods (e.g., days). Over time the vertical width of the
parabola could increase or decrease on a daily basis in response to changes in the cross-sectional return dispersion
among assets (i.e., the variability of returns across all assets at any given time).
Assuming a large number N of assets that is not infinite in number, in theory a return/standard deviation
hyperbola exists (or a return/variance parabola).1 A dashed line is shown at $$E(R_G)$$ that divides the
hyperbola into equal and symmetric halves. This line is known as the axis of symmetry.
For a large number of N securities, the average portfolio return can be written more generally as follows:
$$\begin{aligned} \mu _P = \sum _{j=1}^N w_j E(R_j), \end{aligned}$$ (1.15)
where $$\mu$$ denotes the population mean return, $$w_j$$ is the weight for the jth security,
$$\sum _{j=1}^N w_j =1$$ , and other notation is as before. The variance of portfolio returns for N securities
is:
$$\begin{aligned} \sigma _P^2 = \sum _{j=1}^N w_j^2 \sigma ^2_j + \sum _{j=1}^N \sum _{k \ne j}^N
(1.16)
w_j w_k \rho _{jk}\sigma _j \sigma _k, \end{aligned}$$
where $$Cov (R_j,R_k) = \sigma _{jk} = \rho _{jk}\sigma _j \sigma _k$$. In terms of the covariances of
security returns, the same formula can be alternatively written as:
$$\begin{aligned} \sigma _P^2 = \sum _{j=1}^N \sum _{k = 1}^N w_j w_k Cov (R_j,R_k) = \sum
(1.17)
_{j=1}^N \sum _{k = 1}^N w_j w_k \sigma _{jk}. \end{aligned}$$
In this realistic case of many securities or assets, the computation of the optimal weights to minimize variance
at different feasible expected return levels is considerably more complicated than the two-asset case discussed
above. Appendix B reviews Markowitz’s solution to this optimization problem. In the real world, some problems
can arise in this solution. For example, some assets will get very low weights that do not allow their inclusion in
the portfolio (e.g., the cost of the investment is less than determined by the weight). Another challenge is that the
portfolio weights need to be adjusted over time to maintain efficient portfolios. To do this, rebalancing transactions
costs must be incurred by buying and selling assets. These costs can offset potential diversification benefits. Yet
another issue is that some researchers have documented that the out-of-sample performance of Markowitz
portfolios is not very good. In this regard, Michaud (1989) found that even simple equal-weighted portfolio
returns have outperformed optimally weighted mean-variance efficient portfolios.
What could explain this poor performance? It is well known that estimation errors can be serious in terms of
accurately measuring input values such as expected returns that are proxied by average historical sample returns.
Also, portfolio performance can be diminished due to other issues, including liquidity requirements, management
policy guidelines, and benchmarks to use for performance evaluation.

1.3 Summary
According to Markowitz (1952, 1959), investors follow a two-step portfolio section process. First, they assess the
potential future (ex ante) performance of securities. Second, they choose a portfolio of securities based on
maximizing the discounted values of future returns by using the present value formula. In this formula, expected
cash flows are in the numerator, and the expected rate of return or discount rate is in the denominator. The discount
rate contains the riskless rate plus a risk premium related to the risk of the security and the risk aversion of
investors. The expected return and variance of returns (or risk) can be computed from the probabilities of returns in
different states of the world.
 In a portfolio, the expected return and variance of returns (risk) are affected by the proportionate weights used
for the securities, their individual expected returns and variances of returns, and the correlation of their returns.
The latter correlation results in a covariance term. If two assets have less than perfect correlation, the covariance
term will be reduced, which decreases the variance of returns. Hence, diversification reduces the risk of the
portfolio but not its expected return.
Investors can use ex post return information to compute average returns and the variance of returns over time
for any security. By weighting security holdings, an investor can evaluate the optimal weights that maximize
average returns per unit standard deviation of returns (or variance of returns). Mathematical methods exist to
compute the optimal weights of different securities in a portfolio. Minimum variance portfolios define the mean-
variance investment parabola of Markowitz. The smallest variance portfolio is the minimum variance global
portfolio G. The efficient frontier is the upper boundary of the parabola to the right of G. Minimum variance
portfolios on the lower boundary of the parabola are inefficient. In the real world, some challenges exist in
implementing minimum variance portfolios on the efficient frontier. Extremely small weights for assets, trading
costs, and estimation of expected returns from average historical returns have led to poor out-of-sample portfolio
performance of optimally weighted portfolios. Further work is needed to realize the dream of highly diversified,
efficient portfolios in professional investment management.

Questions
1.
What is the two-step investment process proposed by Markowitz?
2.
Write the general present value formula used to maximize the discounted value of future investment returns.
What are the two components of the discount rate?
3.
Assume that there are s states of nature. Write the formulas for the expected return of security j and the
variance of security j returns. Next assume that there are two securities 1 and 2 in a portfolio. Write the
formulas for the expected return and variance of portfolio returns.
4.
What is the difference between ex ante and ex post returns?
5.
Write the formulas for ex post computation of mean or average returns for security j for time
$$t = 1, \ldots , T$$ periods and the sample variance of returns.
6.
According to Markowitz, what is wrong with maximizing returns of a portfolio as an investment goal? How
can investors use ex post data to evaluate an investment?
7.
How does the correlation between different assets in a portfolio affect the risk of the portfolio?
8.
Figure 1.2 in the text plots the mean-variance parabola of Markowitz. Why is the parabola curved in shape?
What is the minimum variance portfolio? What is the efficient frontier? Are some portfolios on the parabola
inefficient?
9.
In the real world, there are millions of assets. Why have investors had difficulty implementing Markowitz’s
optimization methods to find the optimal weights of assets to hold and therefore hold portfolios on the mean-
variance parabola?

Problems
1.
Table 1.1 contains some computations of the expected rate of return and variance of returns for an asset.
Assume that the probability $$p_s$$ are changed to 0.30, 0.40, and 0.30 in states 1, 2, and 3, respectively.
Recompute the results to create a new Table 1.1.
2.
Table 1.2 computes the average return and variance of returns for two securities. For security 1, change the
returns for years 1 through 5 to the return series 5, 10, 6, 12, and 15. For security 2, change the return series to
8, 9, 8, 5, and 12. Recompute the results to create a new Table 1.2.
3. Table 1.3 computes the portfolio return and risk (in percent) using the data in Table 1.2. Use the data in your
new Table 1.2 from problem 2 to create a new Table 1.3.

Appendix A: Optimal Weights for Two Securities

Here we assume two securities with returns that have correlation $$-1< \rho < 1$$. We want to find the
optimal weights to minimize the portfolio variance. We invest proportion w in security 1 and proportion
$$(1-w)$$ in security 2 that add to one. Using previous formulas, the variance of portfolio returns is defined as:
$$\begin{aligned} \sigma _P^2 = w^2 \sigma ^2_1 + 2 w (1-w) \rho _{12} \sigma _1 \sigma _2 + (1-w)^2
(A1.1)
\sigma ^2_2. \end{aligned}$$
Expanding terms, we have:
$$\begin{aligned} \sigma _P^2 = w^2 \sigma ^2_1 + 2 w \rho _{12} \sigma _1 \sigma _2 - 2 w^2 \rho
(A1.2)
_{12} \sigma _1 \sigma _2 + (1-w)^2 \sigma ^2_2. \end{aligned}$$
Taking the partial derivative, we get:
$$\begin{aligned} \frac{\partial \sigma _P^2}{\partial w} = 2 w \sigma ^2_1 + 2 (1-2w) \rho _{12}
(A1.3)
\sigma _1 \sigma _2 - 2(1-w) \sigma ^2_2. \end{aligned}$$
Setting the above expression to 0 (where variance is no longer reduced for a small increase in w), the optimal
weight for security 1 is:
$$\begin{aligned} w^* = \frac{\sigma ^2_2 - \rho _{12} \sigma _1 \sigma _2}{\sigma ^2_1 - 2\rho _{12}
(A1.4)
\sigma _1 \sigma _2 + \sigma ^2_2}, \end{aligned}$$
and the optimal weight for security 2 equals $$(1-w^*)$$. Note that, if $$\rho = 0$$ is assumed, such that the
covariance is zero, then $$w^*= \frac{\sigma^2_2}{\sigma^2_1 + \sigma^2_2}$$ and
$$(1-w^*) = \frac{\sigma^2_1}{\sigma^2_1 + \sigma^2_2}$$. When covariance is nonzero, it will affect the
optimal weights.

Appendix B: Optimal Weights for Many Assets

This appendix reviews the standard mean-variance portfolio problem. Given the expected returns and matrix of
covariances of returns for N assets, we seek to find the minimum variance parabola based on the set of optimal
asset weights that minimizes the return variance for each feasible portfolio P expected return. Following
Markowitz (1959), a Lagrangian function is used to find the minimum variance portfolio I subject to the
constraints that its expected return equals $$E({R}_I)$$ and constituent asset weights sum to one (i.e.,
associated with Lagrangian multipliers $$\eta$$ and $$\kappa$$):
$$\begin{aligned} \underset{\omega , \eta , \kappa }{\textrm{Min}} L =\frac{1}{2}\omega '\Sigma
(B1.1)
\omega +\eta [E(R_I)-\omega 'E(R_P)]+\kappa (1-\omega 'e), \end{aligned}$$
where $$\omega$$ is the vector of weights (with dimension $$N\times 1$$), e is a vector of ones (with
dimension $$N\times 1$$), $$E({R}_P) = [E({R}_1), E({R}_2), \ldots , E({R}_N)]'$$ is the expected one-
period return vector for N assets (with dimension $$N \times 1$$), and $$\Sigma$$ is the covariance matrix
for N assets (with dimension $$N\times N$$).
To optimize function L, we take the following first derivatives for each asset weight in the vector
$$\omega$$ to obtain $$N+2$$ first-order conditions:
$$\begin{aligned} \frac{\partial L}{\partial \omega } = \Sigma \omega - \eta E(R_P) - \kappa e =
(B1.2)
0\end{aligned}$$

$$\begin{aligned} \frac{\partial L}{\partial \eta } =E(R_P) - \omega ^{\prime } E(R_P) =

(B1.3)
0\end{aligned}$$

$$\begin{aligned} \frac{\partial L}{\partial \kappa } = e - \omega ^{\prime } e = 0. \end{aligned}$$ (B1.4)

From Eq. (B1.2),
$$\begin{aligned} \omega= & {} \Sigma ^{-1} (\eta E(R_P) +\kappa e)\nonumber \\= & {} \eta
(B1.5)
\Sigma ^{-1} E(R_P) + \kappa \Sigma ^{-1} e. \end{aligned}$$
Multiplying Eq. (B1.5) by $$E(R_P)^{\prime }$$:
$$\begin{aligned} E(R_P)^{\prime }\omega = \eta E(R_P)^{\prime } \Sigma ^{-1} E(R_P) + \kappa
(B1.6)
E(R_P)^{\prime } \Sigma ^{-1} e = E(R_I). \end{aligned}$$
Multiplying Eq. (B1.5) by $$e^{\prime }$$:
$$\begin{aligned} e^{\prime } \omega = \eta e^{\prime } \Sigma ^{-1} E(R_P) + \kappa e^{\prime }
(B1.7)
\Sigma ^{-1} e = 1. \end{aligned}$$
Setting
$$\begin{aligned} A\equiv & {} E(R_P)^{\prime } \Sigma ^{-1} e\nonumber \\ B\equiv & {}
E(R_P)^{\prime } \Sigma ^{-1} E(R_P)\nonumber \\ C\equiv & {} e^{\prime } \Sigma ^{-1} e, (B1.8)
\end{aligned}$$
and solving for $$\eta$$ and $$\kappa$$, we obtain the known scalars:
$$\begin{aligned} \eta = \frac{C E(R_I) - A}{D} \end{aligned}$$ (B1.9)

$$\begin{aligned} \kappa = \frac{B - A E(R_I)}{D}, \end{aligned}$$ (B1.10)

where $$D = BC - A^2$$.
Solving for the optimal weights $$\omega ^*$$, we get
$$\begin{aligned} \omega ^*= & {} \frac{B-A E(R_I)}{BC-A^2}\Sigma ^{-1}e
+\frac{CE({R}_I)-A}{BC-A^2}\Sigma ^{-1}E(R_P)\nonumber \\\equiv & {} \phi +\psi E({R}_I), (B1.11)
\end{aligned}$$
where
$$\begin{aligned} \phi\equiv & {} \frac{B\Sigma ^{-1}e-A\Sigma ^{-1}E({R_P})}{BC-
A^2}\nonumber \\ \psi\equiv & {} \frac{C\Sigma ^{-1}E({R_P})-A\Sigma ^{-1}e}{BC-A^2} (B1.12)
\end{aligned}$$
are known $$N\times 1$$ vectors. From these equations, the familiar definition of the variance of efficient
portfolio I is
$$\begin{aligned} \sigma ^2_{I}=\frac{1}{C} + \frac{C\left[ E({R}_{I})-\frac{A}{C}\right] ^2}{BC-
(B1.13)
A^2}. \end{aligned}$$
A wide variety of computer programs have been developed to estimate the above equations. Quadratic programs
are known to provide exact optimal solutions, and other programs estimate approximate solutions. There is an
extensive literature on computer algorithms used to estimate mean-variance optimal portfolios. Programs are
readily available for purchase on the internet.

References
Kolari, J.W., W. Liu, and J. Huang. 2021. A New Capital Asset Pricing Model: Theory and Evidence. New York, NY: Palgrave Macmillan.
[Crossref]

Markowitz, H.M. 1952. Portfolio selection. Journal of Finance 7: 77–91.

Markowitz, H.M. 1959. Portfolio Selection: Efficient Diversification of Investments. New York, NY: Wiley.

Michaud, R.O. 1989. The Markowitz optimization enigma: Is “optimized’’ optimal? Financial Analysts Journal 45: 31–42.
[Crossref]

Footnotes
1 Kolari et al. (2021) observed that the efficient frontier could collapse to the Y-axis in the unrealistic case of infinite number of assets. In the real world,
the total number of assets is large but far from this infinite limit condition.
© The Author(s), under exclusive license to Springer Nature Switzerland
AG 2023
J. W. Kolari, S. PynnönenInvestment Valuation and Asset
Pricinghttps://doi.org/10.1007/978-3-031-16784-3_2

2. Capital Market Conditions

James W. Kolari1 and Seppo Pynnönen2
(1)
Mays Business School, Texas A&M University, College Station, TX, USA
(2)
Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]
Supplementary Information

The online version contains supplementary material available at https://doi.

org/10.1007/978-3-031-16784-3_2

This chapter covers different aspects of capital markets that are essential to
understanding asset pricing. In the previous chapter, we learned that market
participants make investment choices based on the expected returns and
risks of assets. Total risk is measured by the variance or standard deviation
of returns. According to Markowitz, investors should diversify to minimize
risk and, in turn, seek to maximize their returns per unit risk. We begin this
chapter by considering the following market conditions that facilitate asset
pricing by investors: perfect capital markets, efficient markets, risk aversion,
and normally distributed returns. While these market conditions do not
always hold in the real world, they may be sufficiently justified by
regulatory and legal practices in countries that promote the development of
modern capital markets.

Next, we discuss the concept of efficient markets in which asset prices

reflect all available market information. The Random Walk
Hypothesis (RWH) posits that asset prices move randomly over time;
consequently, it is not possible to profit from trading on past patterns of
prices. Contrary to the RWH, anomalous price patterns such as momentum
in asset prices have supported behavioralist views that trends in prices can
occur due to psychological factors.

Another important construct in financial markets is investor risk aversion.

Utility theory is used to define risk aversion and measure the risk
premium demanded by investors on risky assets. An indifference
curve shows how risk-averse investors trade-off expected returns and
variance of returns. Assuming quadratic utility, increasing mean returns
increases investor utility, whereas increasing variance of returns decreases
utility.

Lastly, an important market condition is normally distributed returns.

Different returns need to be defined, including simple returns, log returns,
gross returns, arithmetic mean returns, geometric mean returns, and
portfolio returns. Assuming normally distributed returns, the mean and
variance of simple returns can be defined. Additionally, return distributions
are characterized by the skewness and kurtosis of returns. We demonstrate
return distribution concepts using real-world stock return data.

2.1 Perfect Capital Markets

A perfect capital market is an ideal standard. Some characteristics of this
utopic market are:

rational behavior;

all information is freely available;

no taxes or transactions costs;

 all investors are price takers;

perfect certainty.

Rational Behavior. In financial markets investors rationally prefer more to

less wealth. This behavior is consistent with the nonsatiation principle of
economics, which says that people always prefer more to less goods, money,
stocks, bonds, etc. People are indifferent to these different forms of wealth.
They simply want to increase their wealth. Consistent with Markowitz as
discussed in the previous chapter, rational investors seek to maximize their
wealth per unit risk.

In the real world, however, investors may not always act rationally. For
example, suppose that you have a choice between increasing your
wealth and compromising a friendship or family relation. It is possible that
you would not seek to maximize your personal wealth in this situation. For
personal reasons, you may choose to sacrifice potential wealth gains in favor
of your ethical values. From a broader perspective, financial markets can
behave in irrational ways at times. History is replete with price crashes in
financial markets. Asset prices begin to fall in the market due to some bad
news. As prices fall, panic can break out among investors, which causes
large numbers of investors to sell en masse with severe downward pressure
on prices. Conversely, at times market optimism or irrational exuberance1
can push asset prices up to excessively high levels known as a bubble.
Normally, bubbles are precursors of crashes as they later burst with
devastating consequences to asset prices.

So-called behavioralists argue that human psychology plays an important

role in financial markets. Investors can over- or underreact to market news
and conditions in response to emotions, biases, social preferences,
environmental factors, etc. Do you fear losses more than you enjoy gains
from investments? Do you weight more recent events more heavily than all
events? Do you feel pressure to conform or follow your peers? Do you
follow fixed rules without fully evaluating all available information? Is a
dollar earned by working at a job valued differently by you from a dollar
received as a gift? Is bad news something that you tend to avoid thinking
about? These questions suggest that human psychology can influence our
investment decisions.
Like so many things, rational behavior can be considered as a continuum
from totally rational to totally irrational. Rationality in asset pricing lies
somewhere along this continuum, with a tendency toward rationality. If
irrationality was fairly common, then rational investors would take
advantage of irrational agents. Rational investors would earn abnormal
returns by buying and selling assets from irrational investors, which creates
arbitrage profits. Assuming there are many rational investors, these
arbitrage profits would gradually diminish and eventually disappear as
prices are pushed up or down by arbiteurs and all such profits are exhausted.
Hence, no-arbitrage profits should exist as long as a sufficient number of
investors are rational. This condition is necessary to achieve equilibrium
asset prices. If arbitrage is unrestricted, arbiteurs will buy and sell to reach
some equilibrium market price.2

All Information Is Freely Available. In order for arbitrage to occur,

information needs to be available to investors. If information is freely
available, equilibrium prices can be rapidly attained. Today, due to the
internet, most investors have access to large amounts of free information.3
Even so, search costs still remain for investors seeking information about
firms which can be significant, especially for large, complex firms.

No Taxes and Transactions Costs. A strong assumption in a perfect capital

market is that no taxes and transactions costs exist. Clearly, taxes can affect
the relative attractiveness of different investment choices beyond their return
and risk characteristics. Investors with different tax rates on income and
capital gains would not evaluate prices on a before-tax but rather after-tax
return basis. Hence, an asset could have different after-tax returns and risks
across investors with different tax exposures. Likewise, transactions costs
can vary across investors and change the returns and risks of assets.
Accounting practices, regulatory constraints, institutional trading, and other
market imperfections can alter transactions costs and impede the formation
of equilibrium prices. The question is: Are arbitrage profits consistently
available due to market imperfections such as tax and transactions costs?

In answer to this question, operational efficiency in financial markets is

steadily increasing due to technological advances in computers, software,
and communications. These trends reduce transactions costs and make
markets more perfect than otherwise. Also, at least among big institutional
investors, which dominate a large proportion of investment activities in the
markets nowadays, tax exposures are similar. Although tax and transaction
imperfections exist, their magnitudes are likely insufficient to prevent the
formation of equilibrium prices in capital markets.

All Investors Are Price Takers. No investor should be able to move prices
by buying and selling assets, thereby becoming a price maker. Instead, all
investors should be price takers. Investors have equal and costless access to
information needed to price assets. If an investor did have an ability to set
prices, they could unfairly achieve arbitrage profits. Again, arbitrage profits
do not exist in a perfect capital market.

Perfect Certainty. Lastly, all investors agree about the future expected
returns and risks of assets. This certainty implies homogeneous
beliefs among investors concerning the mean-variance investment parabola
proposed by Markowitz. Without this assumption, no equilibrium prices are
available to construct the investment parabola. In this regard, it should be
noted that, even in uncertain markets, homogeneous expectations are
essential to equilibrium asset prices.

Are capital markets perfect in the real world? Of course, this possibility is
far from reality. Nonetheless, this assumption is a useful simplification in
theoretical derivations of asset pricing models, which we cover in
forthcoming chapters. No model is completely valid due to the violation of
perfect markets. Instead, the validity of models is determined by their
empirical success in helping to explain realized returns and risks in the
financial marketplace. If they do a good job based on actual price data, we
can infer that violations of the perfect capital markets assumption are not so
large as to prevent us from measuring returns and risks with a fair degree of
accuracy.

2.2 Efficient Markets

Efficient markets depend on the existence of a perfect capital market with
no-arbitrage profits. Prices are efficient if they reflect all available
information at any point in time. It is worthwhile to note that a perfect
capital market is a sufficient but not necessary condition for an efficient
market. That is, absolutely perfect capital market conditions are not needed
for the market to operate efficiently, only some reasonable level of such
conditions. How can we determine if sufficient conditions have been met?
According to the efficient market hypothesis (EMH) proposed by Eugene
Fama (1970), who won the Nobel Prize in Economics in 2013, the following
empirical tests of market efficiency can provide insights into the answer to
this question:

weak-form efficiency in which future prices cannot be predicted from

current or past historical prices;

semi-strong-form efficiency in which prices reflect publicly available

information; and

strong-form efficiency in which prices reflect private information.

Weak-form efficiency implies that an investor cannot make profits from

trading on past price movements. Some investors use “technical analysis”
that takes into account past patterns of prices and then seek to trade on these
patterns. In general, most experts would agree that so-called technicians are
unable to earn abnormal trading profits. One reason is that prices tend to
randomly change over time. The Random Walk Hypothesis (RWH) of
Burton Malkiel (2003) says that prices move around in a random manner
much like a random particle in physics. Indeed, tests of prices using physics
models have confirmed this description of price behavior. Prices over time
appear to be independent of past prices.4

A major challenge to RWH is the well-known anomaly of momentum in

asset prices. In a seminal paper in this large field of academic study and
professional practice, Jegadeesh and Titman (1993) showed that abnormal
trading profits can be earned on simple strategies that buy (sell) recent
winners (losers). Because momentum relies upon choosing asset
investments based on recent trends in prices over the past 3-to-12 months, it
violates the EMH. How can such profits be possible?

As mentioned earlier, behavioralists argue that investors are affected by

overconfidence, herding, and other psychological factors that can lead to
overreactions or underreactions to information that cause trends in prices.
Alternatively, momentum profits may arise from risks associated with such
investment portfolios. Here is where Fama’s EMH plays an important role.
If prices are efficiently priced with respect to all risks, no abnormal profit
should exist. By implication, in order to test the EMH, a joint hypothesis is
required: (1) the asset pricing model used to measure risk is correct; and (2)
prices reflect all available information related to risk. This joint hypothesis
creates difficulties for asset pricing models. If the empirical evidence does
not support a particular model, we do not know if the model failed or if
market efficiency failed. Conversely, if evidence did support a model, this
finding would corroborate both the model and market efficiency.

Many tests of publicly available news about companies have shown that
semi-strong efficiency holds in financial markets. Fama showed that prices
react rapidly to news of a stock split by a company. Many event studies of
news announcements about new products, patent approvals, changes in
profits, etc., have confirmed that prices are fairly efficient in response to
most public information. However, there are numerous studies that find
persistent abnormal price movements lasting over a year or more in response
to major corporate actions, including mergers and acquisitions, new issues
of stocks, stock repurchases, dividend initiations, etc. Again,
behavioralists attribute these long-run abnormal prices (and returns) to
psychological factors. Others believe that changing risk explains these long-
run anomalies. If an asset pricing model correctly adjusts for risk, no long-
run abnormal returns or profits are possible. In Chapter 11, we provide
further discussion of event studies, which employ asset pricing models in
their analyses.

Strong-form market efficiency implies that investors cannot profit from

private, inside (or monopolistic) information. In this regard, it is illegal for
managers as insiders to trade on information about the company. If they
knew that a merger with another company was imminent, they cannot
legally trade in the company’s stock until this information is publicly
disclosed. Hence, the market is strong-form inefficient in a regulated
financial world.

In sum, sufficient conditions appear to be present for the existence of

efficient markets in modern financial systems. If market prices are efficient,
we can infer that expected returns reflect all available information. As such,
expected returns are conditional on information about both future cash flows
and risks. As discussed in the previous chapter, Markowitz established the
two-parameter, mean-variance paradigm to measure expected returns and
associated risks. How do individuals make decisions about investment
choices when faced with risk? We turn next to understanding investor
preferences for risk.

2.3 Risk Aversion

Early writers von Neumann and Morgenstern (1944) as well as Friedman
and Savage (1948) considered the problem of an individual that must choose
among investment alternatives with different risks. To solve this problem,
individuals must quantify their preference for one investment over another
using a construct called utility (U). If the expected return of investment A
has more utility than B, and B over C, then we know that they can be ranked
in terms of expected utility as $$U(A)> U(B) > U(C)$$. This rank
takes into account the probability distributions of returns that embody the
riskiness of assets.

Another characteristic of utility is that, for risk-averse investors, it is

concave in shape. As expected utility increases, each incremental increase is
worth less to the individual. In terms of mathematics, the first derivative of
utility is positive (i.e., individuals want more to less utility) but the second
derivative is negative (i.e., each additional unit of utility has less value to the
investor than the previous unit). Risk loving investors would actually have a
positive second derivative, as they enjoy each unit of utility more than the
previous unit. Risk neutral investors have a zero second derivative, such that
each increment of utility is the same as the previous increment. In the real
world, most investor behavior conforms to risk aversion with diminishing
marginal utility.

As an example of risk, Friedman and Savage proposed the following

alternative choice problem. Investment A offers the portfolio combination of
investment return $$I_1$$ with probability $$\alpha$$ (
$$0< \alpha < 1$$) and investment return $$I_2$$ with probability
(1 - $$\alpha$$). Investment B is a riskless return paying certain return
$$I_0$$. Using these probabilities, the expected utility of investment A is:
 $$\begin{aligned} \bar{U}(A) = \alpha U(I_1) + (1 - \alpha ) U(I_2).
\end{aligned}$$
(2.1)
If $$\bar{U}(A)$$ is greater than $$U(I_0)$$, the investor will choose
A over B. Also, if $$\bar{U}(A)$$ is less than or equal to $$U(I_0)$$,
then the investors will prefer B or be indifferent between A and B,
respectively. Now let’s define the actuarial value of A in dollars as:
$$\begin{aligned} \bar{I}(A) = \alpha I_1 + (1 - \alpha ) I_2.
\end{aligned}$$
(2.2)
Assuming that $$\bar{I}(A) = I_0$$, what if the investor prefers A to B?
This preference for risk means that $$\bar{U}(A) > U(I_0)$$, such
that $$\bar{U}(A) - U(I_0)$$ equals the utility gained by the investor in
taking the risk associated with choice A. An investor that does not like risk
will choose a certain investment B with $$\bar{U}(A) < U(I_0)$$.
And, an investor that is indifferent to risky investment A and certain
investment B has $$\bar{U}(A) = U(I_0)$$.
This framework enables an analysis of investment choice under risk. To see
this, suppose that there is some certain return $$I^*$$ with the same
utility as A, i.e., $$\bar{U}(A) = U(I^*)$$. If this certainty
equivalent return $$I^*$$ is less than the actuarial value of A, then the
investor is risk averse and is willing to pay up to $$\bar{I}(A) - I^*$$ as
insurance against the risk of A. However, what if the investor likes risk? In
this case, the investor prefers to gamble and is willing to pay up to
$$I^* - \bar{I}(A)$$ to do so.

Fig. 2.1

Utility analysis of choices involving risk

Based on graphs in Friedman and Savage (1948), Fig. 2.1 depicts the above
utility analysis of alternative investments under risk. The actuarial value of
investment A in assets $$I_1$$ and $$I_2$$ is shown along the straight
line at $$\bar{I}(A)$$. Investment A has utility equal to the certainty
equivalent return $$I^*$$, or $$\bar{U}(A) = U(I^*)$$. Recall also that
investment B has certain return $$I_0$$. The left (right) graph illustrates
risk averse (risk preference) behavior by an investor. To simplify matters,
we adopt the reasonable assumption that certain return $$I_0$$ from
investment B is less than risky return $$\bar{I}(A)$$ from investment A.5

In the risk averse graph on the left, with diminishing marginal utility, if
$$I_0$$ is greater than certainty equivalent return $$I^*$$, the investor
will prefer certain investment B with higher utility. Furthermore, in the
likely case that $$I_0 < \bar{I}(A)$$, the investor would be willing to
pay $$\bar{I}(A) - I_0$$ for this certainty, which is the same as “buying
insurance” to obtain the certain outcome. This payment for risky investment
A is referred to as the risk premium, which is required by the investor for
taking risk. However, if $$I_0 < \bar{I}(A)$$ and also
$$I_0 < I^*$$, investment A is preferred with higher utility. This
framework describes risk-averse behavior among investors.

The graph on the right shows how investors who prefer risk behave. As
discussed above, risk loving behavior is reflected in the utility curve
increasing at an increasing rate and $$I^* > \bar{I}(A)$$. In the likely
case that $$I_0 < \bar{I}(A)$$, higher utility is gained from risky
investment A compared to certain investment B. But unlike risk-
averse behavior, the investor would be willing to pay $$\bar{I}(A) - I_0$$
for risky investment A, i.e., this situation can also be interpreted as “selling
insurance.”

Fig. 2.2

Expected returns and utility for multiple risky investment choices

The previous utility analysis of risk focused on a simple choice between

investments A and B. In the real world, investors are confronted with many
investment choices. For example, Fig. 2.2 provides data for different
investments with alternative payoff gains and losses. Based on probabilities
of gains ( $$\alpha$$) and losses ( $$1 -\alpha$$) for each investment,
expected returns and expected utilities can be computed. The graph
illustrates the diminishing marginal utility of returns as returns increase for a
risk-averse investor.
When evaluating multiple investments, as proposed by Markowitz, risk-
averse investors take into account not only the expected return but the
variance of returns also. The variance (or standard deviation) measures the
dispersion of returns around the expected return (or mean return). High
(low) dispersion implies more (less) negative and positive deviations from
the mean and is associated with higher (lower) risk. Higher expected returns
can be obtained by taking higher risks. Hence, according to Tobin (1958),
investors are indifferent between the utility of various combinations of
expected returns on investments and their variances.

Fig. 2.3

Mean-variance indifference curve of a risk-averse investor

Figure 2.3 shows a graph of Tobin’s mean-variance indifference curves. The

investor is indifferent between all mean-variance points that lie on the
indifference curve. As variance (or risk) increases, risk-averse investors
demand increasingly higher mean returns per unit risk, which gives the
curve its concave upward shape.6 Investors will prefer the indifference
curve 2 to 1 due to the fact that, for a fixed variance (risk), a higher mean
return can be obtained from investments along curve 2 compared to curve 1.
Also, for a fixed mean return, risk-averse investors prefer the lower variance
(or risk) of curve 2 compared to curve 1.

Tobin further proposed that the mean-variance paradigm can be justified by

assuming a quadratic utility function. Given mean returns R and variance of
returns $$R^2$$, the utility function of the investor takes the form:
$$\begin{aligned} U(R) = (1 + b)R + bR^2, \end{aligned}$$
(2.3)
where $$-1< b < 0$$ for a risk-averse investor.7 The first term
$$(1 + b)R$$ implies that utility increases as mean returns increase. The
second term $$bR^2$$ indicates that utility decreases as variance (or risk)
increases (as the b coefficient has a negative sign). Here it must be assumed
that marginal utility is always positive with respect to increasing returns.8

Quadratic utility is an important assumption in the Markowitz two-

parameter, mean-variance framework. Also, as we will see in the next
chapter, quadratic utility is foundational to the Capital Asset Pricing Model
(CAPM) of Treynor (1961, 1962), Sharpe (1964), Lintner (1965), and
Mossin (1966), which assumes that investors obtain equilibrium prices by
taking into account the mean and variance of returns.

2.4 Normally Distributed Returns

Let’s take a closer look at the effects of mean and variance on the
distribution of returns. Before doing so, we briefly discuss some useful
definitions of stock returns. The major reason for analyzing returns rather
than prices is that, while prices are highly persistent and therefore
predictable, returns bear no predictability in efficient markets. Therefore,
returns are the major ingredients when investigating the behavior of
financial markets.

Asset returns are based on price changes between adjacent time periods
$$t-1$$ to t. Using stocks as the asset, simple returns are defined as
$$\begin{aligned} R_t = (P_t + D_t - P_{t-1}) / P_{t-1}, \end{aligned}$$
(2.4)
where $$P_t$$ is the stock price at time t, $$D_t$$ is the dividend paid
on the stock during the period from $$t - 1$$ to t, and $$P_{t-1}$$ is
the stock price at previous time $$t-1$$. Assuming an infinite number of
subperiods within period t, continuously compounded returns can be
computed using log returns as
$$\begin{aligned} r_t = \log (P_t + D_t) - \log (P_{t-1}), \end{aligned}$$
(2.5)
where $$\log$$ is the natural logarithm. Simple and log returns are related
to one another as follows:
$$\begin{aligned} r_t = \log (1 + R_t), \end{aligned}$$
(2.6)
where $$1 + R_t$$ is the gross return, and $$1 + r_t$$ is the gross log
return.

It is important to note that returns are scale free but not unitless, which
means that they are always defined with respect to some time interval, e.g.,
one month. The typical jargon is (for example) “5 percent per month” for
$$R_t = 0.05$$ measured over a one-month period of time.
Assuming no dividends, and using the definition of the gross return
$$1 + R_t = P_t / P_{t-1}$$, we can compute the k-period gross return as
a product of the single-period returns from $$t - k + 1$$ to t as follows:
$$\begin{aligned} 1 + R_{t} (k) & = (1 + R_{{t - k + 1}} ) \times (1
+ R_{{t - k + 2}} ) \times \cdots \times (1 + R_{t} ) \\ & =
\frac{{P_{{t - k + 1}} }}{{P_{{t - k}} }} \times \frac{{P_{{t - k + 2}} }}
{{P_{{t - k + 1}} }} \times \cdots \times \frac{{P_{t} }}{{P_{{t - 1}} }} =
\frac{{P_{t} }}{{P_{{t - k}} }}. \\ \end{aligned}$$
(2.7)
This relation only holds if dividends are ignored. If dividends are paid at
some point within the period, then the single-period returns in the equation
can be adjusted to compute the k-period return. In this case, the equation
implicitly assumes that dividends are reinvested in the stock over time. In
the same fashion, other adjustments such as stock splits or new issues can be
taken into account at the time they occur.
Computing a mean return over some time period k is more complicated. The
arithmetic mean over the whole sample period of $$t = 1, \ldots , T$$ is
defined as
$$\begin{aligned} \bar{R} = \frac{1}{T} \sum _{t = 1}^T R_{t}.
\end{aligned}$$
(2.8)
The arithmetic mean of k period returns from $$t - k + 1$$ to t is
mathematically defined as:
$$\begin{aligned} \bar{R}_t(k) = \frac{1}{k} \sum _{j = 1}^k R_{t - j +
1}. \end{aligned}$$
(2.9)
However, this mean return is always biased upward and therefore is too
high. The main problem is that returns on investment develop in a
compounded manner, not additive. A simple example illustrates the bias. If
you start with $200 and lose 20% in the first period, you have
$$(1 - 0.20)\times \$200 = \$160$$ to invest in the second period. If the
gain in this second period is 30%, the end value becomes
$$(1 + 0.30)\times \$160 = \$208$$, such that the 2-period return of your
investment is $$(208-200) / 200 = 4\%$$, i.e.,
$$(1 - 0.20) \times (1 + 0.30) - 1$$. The additive return of 10% (
$$= -0.20 + 0.30$$) in the arithmetic mean formula is clearly too high and
therefore yields an upward-biased mean return equal to 5% ( $$=10\%/2$$
).
The correct mean return in a multiperiod context is the geometric mean
computed as:
$$\begin{aligned} \bar{R}_g = \left[ \prod _{t = 1}^T(1 + R_t)\right]
^{1/T} - 1 \end{aligned}$$
(2.10)
for the whole sample period, and for a subperiod of k returns from
$$t - j + 1, \ldots , t$$ the geometric mean becomes:
$$\begin{aligned} \bar{R}_{g,t}(k) = \left[ \prod _{j = 1}^k (1 + R_{t - j
+ 1})\right] ^{1/k} - 1. \end{aligned}$$
(2.11)
Note that the whole sample notations are simplifications of the subperiod
notations. That is, for the whole sample $$k = T$$ and therefore for
example $$\bar{R}_{g,t}(k) = \bar{R}_{g,T}(T) = \bar{R}_g$$.

The geometric mean is more complex in mathematical terms but simplifies

matters in the sense that, given the geometric mean, the k-period gross
return becomes $$1 + R_t(k) = [1 + \bar{R}_{g,t}(k)]^k$$. Thus, in the
simple example above, the correct mean return is
$$[(1 - 0.2)\times (1 + 0.3)]^{1/2} - 1 = 1.98\%$$ (rather than the
arithmetic mean of $$(-0.2 + 0.3) / 2 = 5\%$$). The 2-period return is then
$$(1 + 0.0198)^2 - 1 = 4\%$$ which matches the correct 2-period return
computed above.

Using log returns, again called continuously compounded returns, the

logarithm of $$1 + R_{t}(k)$$ transforms the product of the gross
returns of securities to a sum as follows:
$$\begin{aligned} r_{t}(k) = \log [1 + R_t(k)] = \sum _{j = 1}^k\log (1 +
R_{t - j + 1}) = \sum _{j = 1}^k r_{t - j + 1} \end{aligned}$$
(2.12)
with $$r_{t - j + 1} = \log (1 + R_{t - j + 1})$$. Therefore the arithmetic
mean
$$\begin{aligned} \bar{r}_t(k) = \frac{1}{k} \sum _{j = 1}^k r_{t - j +
1} \end{aligned}$$
(2.13)
correctly measures the average multiperiod log return also. Moreover, it is
useful to observe that $$\bar{R}_{g,t}(k) = e^{\bar{r}_t(k)} - 1$$ and
 $$1 + R_t(k) = (1 + \bar{R}_{g,t}(k))^k = e^{k\bar{r}_t(k)}$$, where
$$e \approx 2.718$$ is the Neper’s or Napierian number after the Scottish
mathematician John Napier (e is also known as Euler’s number after the
Swiss mathematician Leonard Euler).

As noted earlier, while returns are scale free, they are not unit free. That is,
they depend on the time unit. Therefore, when discussing returns, it is
important to indicate the unit, whether it is daily, weekly, monthly, yearly,
etc. In particular, when making comparisons, it is important to transform
returns to the same time unit. It is not useful to annualize single daily
returns. A simple example demonstrates the problem. Assume a 1% return
on some day. If annualized by assuming 365 days, one gets
$$(1 + 0.01)^{365} - 1 \approx 36.78$$ or 3,678%, implying that the
transformation explodes the numbers. However, average daily returns can be
annualized as averaging smooths single day variation. The k-period
geometric mean in Eq. (2.11) can be used for the annualization. Also, even
though it is biased, sometimes the k-period arithmetic mean in Eq. (2.9) is
used.

Regarding simple versus log returns, in high frequency data based on daily
or weekly returns, the numerical value differences between simple
returns and log returns are usually negligible as demonstrated shortly.
However, there are fundamental differences between these return concepts.
One difference is that the simple return is bounded below by $$-1$$ when
the stock price falls to zero. In this case, log returns approach minus infinity.
Therefore, while simple returns can take values from minus one upward, log
returns can be all real values.

A more important difference between these returns is the way they

aggregate with multiple assets in a portfolio. We previously observed that,
unlike log returns, summing simple returns over time does not aggregate
correctly to total returns. The opposite is true when summing cross-
sectionally stocks into portfolios at a point in time. Portfolio returns can be
obtained as weighted sums of the simple returns, while log returns of a
portfolio cannot be obtained as a weighted sum of log returns of multiple
stocks. For this reason, it is common to use simple returns when the cross-
section of assets is studied in some period of time and log returns when the
temporal behavior of returns over time is under study.
In discussing statistical properties, it is convenient to start with the log
return because of its summation property shown in Eq. (2.12). We can think
of each one-period return, $$r_t$$, as the sum of subperiods between
times $$t-1$$ and t—for example, days can be divided into hours or even
minutes. Thus, dividing the time interval from $$t-1$$ to t into n
(nonoverlapping equal length) subperiods, or
$$t-1 = t_0< t_1, \ldots , < t_n = t$$, and defining log returns as
$$r_{t_i} = \log (P_{t_i}) - \log (P_{t_{i-1}})$$, we have
$$\begin{aligned} r_t = \sum _{i = 1}^n r_{t_i}. \end{aligned}$$
(2.14)
If markets are efficient, intra-period returns should be approximately
independent. Under some fairly general conditions, the Central Limit
Theorem (CLT) implies that, as n grows (i.e., we take finer and finer
subperiods), the distribution of the sum approaches that of a normal
distribution. That is, if the assumptions hold, the CLT predicts that
$$r_t$$ is normally distributed. Furthermore, the properties of normally
distributed random variables imply that, if we take an exponential
transformation (anti-log transformation) of a normally distributed random
variable, the transformed random variable follows the lognormal
distribution. Therefore, because the gross return $$1 + R_t = e^{r_t}$$,
the theory predicts that $$1 + R_t$$ follows a lognormal distribution.

Unfortunately, due to complexities in the real world, the normal

distribution has proven often to be a fairly poor approximation of returns,
particularly for the distribution of high frequency returns (intra-day and
daily returns). Stock returns appear to have the feature that, while they are
serially virtually uncorrelated over time, squared returns and absolute values
of returns are highly serially correlated from period to period. This breaks
the independence assumption of the CLT and disallows the normality
implication of the theorem. This issue is a concrete example of the fact that
zero correlation and statistical independence are not the same. It is an
important fact that stock return distributions have far fatter tails than those
of the normal distribution.

In spite of these drawbacks, the means and variances of returns that

characterize the (multivariate) normal distribution are the key parameters in
asset pricing. Let’s assume for the moment that the log return, or
$$r_{it}$$, for stock i is normally distributed, which can be denoted as
 $$\begin{aligned} r_{it} \sim N(\mu _i, \sigma _i^2), \end{aligned}$$
(2.15)
where $$\mu _i = E(r_{it})$$ and $$\sigma _i^2 = Var [r_{it}]$$ are,
respectively, the mean and variance of returns. Then the relation between the
normal distribution and the lognormal distribution implies that the mean and
the variance of the simple return $$R_{it}$$ are
$$\begin{aligned} E(R_{it}) = e^{\mu _i + \frac{1}{2} \sigma _i^2} - 1
\end{aligned}$$
(2.16)
and
$$\begin{aligned} Var (R_{it}) = e^{2\mu _i + \sigma _i^2}\left(
e^{\sigma _i^2} - 1\right) . \end{aligned}$$
(2.17)
Obviously, the corresponding lognormal formulas would be much more
complicated. However, for high frequency daily returns, the Taylor
expansion:9
$$\begin{aligned} R_{it} = e^{r_{it}} - 1 = r_{it} + \frac{1}{2} r_{it}^2
+ \text {higher order terms}, \end{aligned}$$
(2.18)
shows that
$$\begin{aligned} R_{it} \approx r_{it}. \end{aligned}$$
Therefore, for daily returns, the two return concepts do not differ materially.
A rule of thumb is that the above approximation is precise up to a 10%
return. Using Eq. (2.18), for a 10% return the error is only about
$$\frac{1}{2}(.01)^2 \approx .005$$, or about half of a percentage point.
In addition to the mean and variance, two additional popular measures
utilized in summarizing return distributions are skewness and
kurtosis defined, respectively, as
$$\begin{aligned} \mu _3 = E\left( \frac{r_t - \mu }{\sigma }\right) ^3
\end{aligned}$$
(2.19)
and
$$\begin{aligned} \mu _4 = E\left( \frac{r_t - \mu }{\sigma }\right) ^4.
\end{aligned}$$
(2.20)
Skewness measures the asymmetry of the distribution, whereas
kurtosis measures the peakedness of the distribution or fatness of the tails.
Because the normal distribution is symmetric, its skewness is zero, i.e.,
$$\mu _3 = 0$$. The kurtosis of the normal distribution is 3. Therefore, it
is a common practice to use excess kurtosis defined as the difference
$$\mu _4 - 3$$, rather than the original kurtosis.
Given T return observations, the sample counterparts for these population
measures are the sample skewness
$$\begin{aligned} \text {sk} = \frac{1}{T - 1} \sum _{t = 1}^T\left(
\frac{r_t - \bar{r}}{s}\right) ^3 \end{aligned}$$
(2.21)
and sample kurtosis
$$\begin{aligned} \text {ku} = \frac{1}{T - 1} \sum _{t = 1}^T\left(
\frac{r_t - \bar{r}}{s}\right) ^4, \end{aligned}$$
(2.22)
where
$$\begin{aligned} \bar{r} = \frac{1}{T} \sum _{t = 1}^T r_t
\end{aligned}$$
(2.23)
is the sample mean of the returns, and s is the return standard deviation,
which is the square root of the sample variance
$$\begin{aligned} s^2 = \frac{1}{T - 1}\sum _{t = 1}^T(r_t - \bar{r})^2.
\end{aligned}$$
(2.24)
If the returns are normally distributed, the skewness and kurtosis are
normally distributed in large samples with respective means 0 and 3 and
variances 6/T and 24/T. Consequently, these features can be utilized in
testing for the normality of a return distribution. Empirically speaking, it has
been observed that the skewness of daily returns of individual stocks is close
to zero and slightly negative for stock indices based on a portfolio of stocks
(e.g., the S &P 500 index).

The excess kurtosis, or $$\text {ku} - 3$$, tends to be positive for both
individual stocks and stock indices, which indicates fatter tails with a more
peaked mean than the normal distribution. Therefore, various alternatives,
such as stable distributions, have been proposed. However, the popularity of
these alternatives has diminished today due to some counterintuitive
properties (e.g., increasing variance with the sample size). On the other
hand, some fatter tailed distributions, like the t-distribution and generalized
error distributions of which the normal distribution is a special case, have
gained popularity over time. In spite of these issues, normal and lognormal
distributions continue to play a central role in theoretical financial
economics. Hence, the empirical challenge is to assess how critical the
normality assumption is for the predictions of theory.

The following empirical example illustrates the concepts and issues

discussed above. We use daily, weekly, and monthly price data
downloadable from Yahoo Finance10 for Microsoft and two stock market
indices, including the S &P 500 Index for the U.S. markets and the DAX
Index for German markets. Our sample time period is from January 1, 2010
until December 31, 2020.
Table 2.1

Sample statistics for S &P 500 index, DAX index, and Microsoft
Corporation: daily, weekly, and monthly returns (in percent)

Daily Weekly Monthly

S &P S &P S &P

MS DAX MS DAX MS DAX
500 500 500

Mean 0.09 0.05 0.04 0.44 0.22 0.19 1.89 1.01 0.77

Median 0.07 0.07 0.08 0.52 0.31 0.42 2.23 1.50 0.71

Volatility 25.39 17.50 20.55 22.69 13.90 20.84 21.33 14.07 17.76

Skewness 0.04 −0.58 −0.38 −0.25 −0.49 −0.82 0.03 −0.31 −0.50

Kurtosis 9.91 15.13 7.21 2.89 1.84 5.69 0.34 0.93 1.65
 Daily Weekly Monthly

S &P S &P S &P

MS DAX MS DAX MS DAX
500 500 500

Min −14.74 −11.98 −12.24 −14.37 −7.19 −20.01 −15.52 −12.51 −19.19

Max 14.22 9.38 10.98 15.02 7.39 10.91 19.63 12.68 15.01

N 2,768 2,768 2,785 574 526 574 132 132 132

Notes Mean and median are percentage returns per indicated period, and
volatility is annualized. Kurtosis is the excess kurtosis, or
$$\text {ku} - 3$$, with $$\text {ku}$$ defined in Eq. (2.22).

Table 2.1 supports the general observation of slight negative skewness of

stock indices (S &P 500 and DAX), which indicates that the left-hand-side
tail is longer than the right-hand-side tail. Using the large-sample test
statistic defined as $$z = \text {sk}\cdot \sqrt{T/6}$$ for testing
symmetry (i.e., whether the population symmetry is zero), in unreported
results, both index skewness measures appear highly statistically significant
for different return frequencies. For individual stocks (Microsoft), skewness
can vary for different sampling frequencies, with both positive and negative
skewness shown in the table. In this regard, the weekly returns are
significantly negatively skewed. These features are observable from the
means and medians in Table 2.1. A positively skewed distribution has a
mean greater than the median, and vice versa for negatively skewed
distributions. Volatilities are annualized in percentage terms. Index
volatilities are generally lower than those of individual stocks due to the
diversification effect discussed in Chapter 1. It is interesting to note that,
while the mean return of the S &P 500 is greater than that of the DAX,
volatility (risk) is lower for the S &P 500, which means that the reward-to-
risk ratio is higher for the S &P 500 over the period 2010 to 2020.

Fig. 2.4

Empirical return distributions for Microsoft, U.S. S &P 500 index,

and German Dax index using daily, weekly, and monthly returns

Figure 2.4 shows the return frequency distributions (in blue) for Microsoft
as well as U.S. S &P 500 and German DAX stock price indexes.
Distributions are shown using daily, weekly, and monthly return
frequencies. These empirical distributions are compared to the imposed
normal curves (in red) calibrated by respective sample means and sample
variances. The peakedness of the empirical distributions illustrates
graphically the kurtosis feature, which is characteristic of both individual
stocks and indexes (portfolios). Fat tails, even though not that obvious from
the graphs, are well-recognized features associated with stock returns due to
high kurtosis. The mild skewness of the return distributions is not obvious
from the graphs. Scanning down the rows of graphs reveals that, as return
frequency decreases from daily to monthly returns, the normal
approximation improves.

Finally, we should note that the above sample statistics and graphs are
produced using simple returns (in percentages). We have left as an exercise
for the student to reproduce these results using log returns and compare if
there are material differences.

2.5 Summary
A number of capital market conditions are foundational to asset pricing.
Perfect capital markets assume rational investor behavior, all information is
freely available, no taxes and transactions costs, all investors are price
takers, and perfect certainty. The efficient market hypothesis (EMH) by
Nobel Laureate Eugene Fama (1970) posits that prices reflect all available
information. Weak-form, semi-strong form, and strong-form
efficiency describe different types of market efficiency. In an efficient
market, the random walk hypothesis (RWH) of Malkiel (2003) argues that
prices move randomly over time and therefore are independent of past
prices. Behavioralists counter that price trends can occur due to investor
psychology which can cause over- or underreactions to information in the
market.

Investors are assumed to be risk averse in capital markets. Research by von

Neumann and Morgenstern (1944) and Friedman and Savage (1948)
employed utility theory to study risky investment choices. If the expected
utility of a risky investment exceeds the utility of a certain return, assuming
their actuarial values is the same, a risk-averse investor will purchase the
risky investment. In the more likely case that the actuarial value of the risky
investment is greater than that of the certain return, a risk premium is paid
by the investor. In a graph of utility (Y-axis) and returns (X-axis), the utility
curve increases at a decreasing rate for risk-averse investors, who obtain less
utility per unit of return as returns increase. For risk loving behavior, the
utility curve increases at an increasing rate as utility per unit risk increases
for each incremental return. In a portfolio context with multiple investments,
Tobin (1958) defined an indifference curve with trade-offs between mean
returns and the variance of returns. Assuming quadratic utility, investors’
utility is increasing in mean returns and decreasing in the variance of
returns.

Stock returns can be defined in a variety of ways. Simple returns take into
account dividends and capital gains on a stock for a discrete time period.
Log returns provide a continuously compounded return. The log of one plus
the simple return (or gross simple return) equals the log return. An
arithmetic mean takes the average of simple returns over a number of time
periods. Because the arithmetic mean return has an upward bias, the
geometric mean is used to compute an average multiperiod return.
Alternatively, the simple arithmetic mean of log returns over a number of
time periods can be used to measure the average multiperiod return. For
high frequency data, such as daily or weekly returns, simple returns, and log
returns are similar to one another, at least up to about a 10% return. When
aggregating returns in a portfolio, simple returns should be used in cross-
sectional analyses of assets at a point in time, and log returns should be used
in time-series analyses of returns over multiple periods. Using log
returns defined as the sum of many subperiod log returns, the Central Limit
Theorem (CLT) says that returns are normally distributed. However, in the
real world, stock returns are not normally distributed; instead, they tend to
have some degree of asymmetry (i.e., skewness) and fatter tails (i.e.,
kurtosis). Mean, variance, skewness, and kurtosis measures are used to
characterize return distributions. Individual stocks normally have higher
variances of returns than portfolios due to the diversification benefit in
portfolios. As the return frequency decreases from daily to monthly returns,
the distribution of stock returns becomes closer to a normal curve.

Questions
1. 1.

In a perfect capital market, rational behavior dictates that investors

prefer more to less wealth. Give some examples of irrational market
behavior. Why might investors act irrationally?

2. 2.

Give three characteristics of a perfect capital market and briefly discuss

them.

3. 3.

Under the efficient markets hypothesis (EMH), what are the three types
of market efficiency? Briefly describe them.

4. 4.

What is the Random Walk Hypothesis (RWH)? Is there any evidence

against this hypothesis?

5. 5.
 What is a certainty equivalent return of a risky investment?

6. 6.

What is a risk premium on an investment?

7. 7.

Figure 2.2 gives an example of a utility curve for a risky investment.

For this risky investment, expected utility is on the Y-axis, and
expected returns are on the X-axis. What is the main concept conveyed
by this graph?

8. 8.

When risk-averse investors evaluate multiple investments in a portfolio

context, what do investors use to evaluate investments? Also, what is
an indifference curve in this respect? Is this curve related to investor
utility? If so, write a utility function that is relevant to the indifference
curve.

9. 9.

Write the simple return for a stock. Also, write the continuously
compounded rate of return. How are these two returns related to one
another?

10. 10.

Are simple returns and log returns similar to one another? In a

portfolio, which type of return should be used to compute the cross-
sectional average return at a point in time? Which one should be used
to compute returns over time in a time-series analysis?
11. 11.

How can the Central Limit Theorem (CLT) be applied to explain why
log returns approach a normal distribution?

12. 12.

Suppose that the simple returns of stocks are normally distributed.

What can you say about the distribution of the return of a portfolio of
these stocks?

Problems
1. 1.

Assume that investment A offers the portfolio combination of

investment return $$I_1$$ with probability $$\alpha$$ (
$$0< \alpha < 1$$) and investment return $$I_2$$ with
probability (1 - $$\alpha$$). What is the expected utility of
investment A, or $$\bar{U}(A)$$? What is the actuarial value of A in
dollars, or $$\bar{I}(A)$$?

2. 2.

Suppose that there is some certain return $$I^*$$ with the same
utility as A, i.e., $$\bar{U}(A) = U(I^*)$$. Also, assume that this
certainty equivalent return $$I^*$$ is less than the actuarial value of
A. What does this tell you about the investor? How much is the investor
willing to pay as insurance against the risk of A?

3. 3.
 Draw a graph of the utility curve showing risk-averse investor behavior
with respect to the risky investment opportunity in problem 1 above.
How would this graph change for risk loving investors?

4. 4.

How would you compute the k-period gross return

for an investment from

to t? Assume no dividends are paid.

Write the general formulas using gross return

. Also,

compute assuming that

$$(1+R_1)=100/90$$, $$(1+R_2)=105/100$$, and

(note $$t = k = 3$$ here).

5. 5.

How can we compute the mean return for an investment over some
time period k? Show the formulas for the arithmetic mean and
geometric mean. Utilizing the gross returns in problem 4, i.e,
$$(1 + R_1) = 1.1111$$, $$(1+R_2) = 1.0500$$, and
 $$(1+R_3) = 1.0476$$, compute the arithmetic and geometric mean
returns. Which mean return should you use and why?

6. 6.

How would we compute the gross return $$1 + R_t(k)$$ over some
time period k using log returns? In this case, what is the simple
arithmetic mean of these gross returns? Is the average of this
multiperiod correct (unbiased)? Assuming that $$(1+R_1)= 1.1111$$
, $$(1+R_2)=1.0500$$, and $$(1+R_3)=1.0476$$, compute these
returns. What is the formula for the geometric mean return?

7. 7.

Consider the monthly returns of a stock. Assume that the returns are
normally distributed with mean $$\mu = 1\%$$ and standard
deviation $$\sigma = 9\%$$ (per month). How often would you
expect to see a monthly return that exceeds $$10\%$$? How about
$$20\%$$.

8. 8.

Download Microsoft stock prices in addition to S &P 500 and German

DAX index values for the sample period from January 1, 2010 until
December 31, 2020 and reproduce the sample statistics in Table 2.1
using log returns. Are there material differences?

References

1. Fama, E.F. 1970. Efficient capital markets: A review of theory and

empirical work. Journal of Finance 35: 383–417.Crossref
2. Fama, E.F., and M.H. Miller. 1972. The Theory of Finance. Hinsdale,
IL: Dryden Press.
 3. Friedman, M., and L. Savage. 1948. The utility analyses of choices
involving risk. Journal of Political Economy 56: 279–304.Crossref
4. Jegadeesh, N., and S. Titman. 1993. Returns to buying winners and
selling losers: Implications for stock market efficiency. Journal of
Finance 48: 65–91.Crossref
5. Lintner, J. 1965. The valuation of risk assets and the selection of risky
investments in stock portfolios and capital budgets. Review of
Economics and Statistics 47: 13–37.Crossref
6. Malkiel, B.G. 2003. A Random Walk Down Wall Street: The Time-
Tested Strategy for Successful Investing. New York, NY: W.W. Norton.
7. Markowitz, H.M. 1952. Portfolio selection. Journal of Finance 7: 77–
91.
8. Markowitz, H.M. 1959. Portfolio Selection: Efficient Diversification of
Investments. New York, NY: Wiley.
9. Mossin, J. 1966. Equilibrium in a capital asset market. Econometrica
34: 768–783.Crossref
10. Sharpe, W.F. 1964. Capital asset prices: A theory of market equilibrium
under conditions of risk. Journal of Finance 19: 425–442.
11. Shiller, R.J. 2000. Irrational Exuberance. Princeton, NJ: Princeton
University Press.
12. Tobin, J. 1958. Liquidity preference as a behavior toward risk. Review
of Economic Studies 25: 65–86.Crossref
13. Treynor, J.L. 1961. Market value, time, and risk. Unpublished
manuscript.
14. Treynor, J.L. 1962. Toward a theory of market value of risky assets.
Unpublished manuscript.
15. von Neumann, J., and O. Morgenstern. 1944. Theory of Games and
Economic Behavior. Princeton, NJ: Princeton University Press.

Footnotes
1

See Shiller (2000), who titled his book “Irrational Exuberance” after a
warning by Federal Reserve Chairman Alan Greenspan of a possible market
bubble in 1996.

2
It is possible that limits to arbitrage can occur due to constraints on traders
arising from institutional or structural restrictions in the financial market.
For example, an investor that borrowed money to buy an asset could face a
margin call to cover losses on the asset that are short run in nature due to a
temporary market panic. These sudden losses could prevent arbitrage gains
from being realized such that equilibrium prices cannot be achieved.

As discussed shortly in the context of strong-form efficiency, private or

inside information of managers of companies exists, but managers cannot
legally trade on this information. Also, securities firms cannot pay firms for
inside information that is not publicly available.

According to Fama and Miller (1972, p. 339, footnote 19), further RWH
assumptions are that prices are identically distributed and that expected
price changes can be nonzero due to drift.

See Friedman and Savage for the case in which $$I_0 > \bar{I}(A)$$.

Risk lovers would have a concave downward-shaped indifference curve.

For a risk lover, we have $$0< b < 1$$.

8
As shown by Tobin (1958, p. 76), this assumption limits the range of mean
returns R.

See Wikipedia on the internet for an excellent review of the Taylor series.
This mathematical method enables the evaluation of a general function with
incremental accuracy by a polynomial of finite order. Often even the first-
order approximation provides a sufficiently accurate approximation for
practical purposes.

10

https://finance.yahoo.com.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_3

3. Capital Asset Pricing Model (CAPM)

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_3

This chapter covers the now famous Capital Asset Pricing Model
(CAPM) as proposed by William Sharpe (1964), for which he was awarded
the Nobel Prize in Economics in 1990. Other authors credited with
contributions to the CAPM are Jack Treynor (1961, 1962), John
Lintner (1965), and Jan Mossin (1966). The CAPM builds upon the mean-
variance paradigm of Harry Markowitz (1952, 1959) and risk-aversion
concepts of Tobin (1958). Interestingly, Sharpe was a Ph.D. student
working closely with Markowitz at University of California at Los Angeles
(UCLA), who shared the Nobel Prize with him along with Merton Miller.
The CAPM represents a landmark that established the study of asset
pricing. Up until that time, finance as an academic and professional
discipline was grounded primarily in accounting principles, present value of
cash flow mathematics, and descriptive risk measures. The CAPM changed
everything. Combining portfolio theory in Chapter 1 with capital market
conditions in Chapter 2, including perfect capital markets, efficient markets,
risk aversion, and mathematical return concepts, the CAPM provided the
first equilibrium asset pricing model. It marked the beginning of a
revolution that elevated the academic and professional field of finance
around the world.

3.1 The Present Value of Future Cash Flows

The present value formula is the workhorse of valuation. The current price
of an asset P at time 0 is determined as follows:
$$\begin{aligned} P_0 =\sum _{t=1}^T \frac{E(CF_t)}
(3.1)
{(1+E(R))^t}, \end{aligned}$$
where $$E(CF_t)$$ are expected cash flows generated by the asset over
time $$t = 1, \ldots , T$$, and $$E(R)$$ is the expected discount rate.
Expected values of these variables are ex ante risky in the sense that they
occur in the future and therefore are not known with certainty. Expected
future cash flows in the numerator can be estimated from accounting,
finance, and operating information about the firm. What were its past sales
revenues, finance costs, and operating costs? Based on the firm’s products
and services, industry competition, and the general economy, what are the
current projections for the cash flows or profits of the firm?
Nowadays many researchers express the present value formula in
single-period t as:
$$\begin{aligned} P_t =E_t(m_{t+1} x_{t+1}), \end{aligned}$$ (3.2)
where $$P_t$$ is the beginning period price, $$x_{t+1}$$
represents cash flows or payoffs in the period from t to $$t+1$$ equal
to $$1+R_{t+1}$$, $$E_t$$ is the expectations operator
conditional on market information at time t, and $$m_{t+1}$$ is the
stochastic discount factor1 (SDF) that discounts expected cash flows to
present value prices. The SDF is defined as
$$m_{t+1} = 1/(1+R_{t+1})$$. Assuming future cash flows are risky,
the discount rate R is random or stochastic. If future cash flows were
certain, then $$m_{t+1} = 1/(1+R_f)$$ or the riskless discount factor.
The above equation can be written in terms of returns by defining
$$R_{t+1} = x_{t+1}/P_t$$ such that
$$\begin{aligned} E_t[m_{t+1}(1+R_{t+1})] = 1. \end{aligned}$$ (3.3)
This value equals 1 due to the law of one price for any asset. Accordingly,
no possibility of arbitrage opportunities exists. As discussed in Chapter 2,
arbitrage occurs when investors buy and sell the same asset at different
prices to make trading profits. Of course, as they buy and sell, the trading
profits are gradually eliminated to cause one price for the assets. This
process assumes no frictions in the market that might prevent investor
arbitrage (e.g., a brokerage firm that charges investors a fee for trading).
Equations (3.2) and (3.3) are the most general forms of asset pricing
models. All models are special cases of these two equations, which fall
under the general heading of “m-talk.”2
The discount factor m in the present value formula is central to asset
pricing. To estimate the expected discount rate in m, it is necessary to
understand the required rate of return of investors in the firm. This return on
investment depends crucially on both the risks of the asset itself and
investor aversion to risks. As discussed in Chapter 1, these two risk
dimensions explain the risk premium equal to the difference between the
risky expected rate of return for the ith asset and riskless rate, or
$$E(R_i) - R_{f}$$.
To estimate discount rate $$E(R_i)$$ investors need to
incorporate all potential risks affecting the price of the asset. For example,
credit risk is related to the probability of bankruptcy of the firm. Liquidity
risk takes into account the ability to buy and sell the asset with little change
in price. Earnings risk considers the ability of the firm to generate profits in
the future. Numerous other risks include (for example) management ability,
legal issues, government regulation, technological change, and agency costs
arising from differences in the objectives of management and shareholders.
As the sum of these and other risks increases, the rate of return required by
investors will increase. A number of questions naturally arise. How should
we measure each of these many different risks? Is there a way to rank the
risks from low to high? Can we combine these risks to estimate overall
asset risk? To complicate matters further, expected future cash flows and
expected discount rates change over time and cause the price of the asset to
adjust or fluctuate in response to these changes.
Another way to estimate $$E(R_t)$$ is by observing historical
prices. As defined in the previous chapter, the ex post rate of return in some
period t is:
$$\begin{aligned} R_t =\frac{P_t-P_{t-1}}{P_{t-1}},
(3.4)
\end{aligned}$$
where no interest, dividends, or other payments to investors are assumed.
The numerator gives the capital gain on the asset. A series of $$R_t$$
values can be computed over time to yield an estimate of the mean rate of
return. The mean return could be used in the present value formula (3.1) as
a proxy for $$E(R_t)$$.
This ex post rate is simple to estimate and provides an historical
estimate of $$E(R_t)$$ but has a major drawback. Namely, it is
backward looking and therefore does not consider future expected rates of
return. Is the current price of the asset too high or low relative to the future
rate of return? An investor could buy an asset that is overvalued with
subsequent losses in its price over time. Likewise, an asset might be sold
that is undervalued and then its price increases causing opportunity costs of
lost gains. How can we determine if an asset is over- or undervalued at any
point in time?
In forthcoming sections, we introduce the now famous Capital Asset
Pricing Model (CAPM) of Treynor (1961, 1962), Sharpe (1964),
Lintner (1965), and Mossin (1966). The CAPM proposes that, once the risk
of an asset is estimated, ex ante rates of return can be determined. An
overall measure of asset risk is derived relative to the market as a whole. In
this sense, asset prices are relative to their market risks, which are jointly
formed in equilibrium. To measure market risk, the CAPM relies upon the
portfolio theory of Markowitz (1952, 1959), who defined the efficient
frontier of all asset returns in the market. Let’s see how Sharpe and other
CAPM authors were able to use modern portfolio theory to solve the
problem of estimating the expected rate of return and market risk of assets.

3.2 Market Equilibrium Assumptions

Sharpe (1964) observed that previous classroom discussions of capital asset
prices lacked a theoretical foundation. Consequently, he sought to develop
an equilibrium market framework for asset pricing. Some idealistic
assumptions were needed to simplify matters. For market equilibrium to be
reached, he assumed: (1) perfect competition exists in capital markets; (2)
investors have homogeneous expectations concerning the
probability distributions of returns on risky assets; (3) investors seek to
maximize the expected utility of wealth; (4) risk-averse investors have
quadratic utility functions; (5) investors can borrow and lend at the riskless
rate; and (6) investors are not allowed to short sell assets in forming
portfolios (i.e., only long asset positions are allowed).
As a starting point, he proposed the Capital Market Line (CML) as
shown in Fig. 3.1. Equilibrium asset prices trade-off the expected rates of
return of assets against their risks. The risk here is measured by the standard
deviation of returns, or total risk. If there is zero risk, the investor earns the
riskless rate $$R_f$$, which is based on the price of time or pure
3
interest rate. Given some expected rate of return per unit risk, under
conditions of risk, the investor demands increasing risk premiums as risk
increases.

Fig. 3.1 Equilibrium relationship between expected rates of return and risk

Sharpe noted that Tobin (1958) used Markowitz’s portfolio theory to

propose a two-step investment process. First, the investor chooses a unique
optimal combination of risk assets in a portfolio. Second, a separate choice
is the allocation of funds to the risky portfolio and a riskless asset. This
Separation Theorem set the stage for Sharpe’s market equilibrium theory of
asset prices under conditions of risk. While preparing his paper, Sharpe
became aware of a similar equilibrium theory developed by Treynor4 that
no doubt influenced his work.

Fig. 3.2 Combining the efficient frontier and indifference curves

Combining Markowitz’s efficient frontier (see Fig. 1.4 in Chapter 1) and

Tobin’s mean-variance indifference curves (see Fig. 2.3 in Chapter 2),
Fig. 3.2 shows that the investor prefers optimal portfolio P. This portfolio
lies at the tangency between indifference curve 3 and the efficient frontier.
Any other portfolio will lower the expected utility of investors (e.g.,
portfolios intersecting indifference curves 1 and 2). The investor maximizes
expected utility at P in mean return and standard deviation return space.
Assuming the investor commits their initial wealth ( $$W_i$$) to
portfolio P, Sharpe concluded that the investor would earn in one period t:
1.
expected rate of return equal to $$E(R_P) = (W_t - W_i)/W_i$$; and
2.
expected utility $$E(U) = g[E(R_P), \sigma (R_P)]$$, where
$$\sigma (R_P)$$ denotes the standard deviation of portfolio P
returns.
Following Tobin’s Separation theorem, the investor initially locates the
efficient frontier and then selects one efficient portfolio with the highest
expected utility.

3.3 Investment Choice with a Riskless Asset

Extending previous authors, Sharpe made a logical leap by introducing the
existence of a riskless asset. Simplifying matters, he assumes a one-period
model, such that we can drop the time subscript t in the notation. Investors
can allocate their wealth among both risky and riskless assets. Using risky
efficient portfolio P from Fig. 3.2, the expected return on this portfolio
combination (C) is:
$$\begin{aligned} E(R_C) = a R_f + (1- a) E(R_P), \end{aligned}$$ (3.5)
where $$R_f$$ is the constant riskless rate for asset f held in wealth
proportion a, and $$E(R_P)$$ is the expected return on portfolio P held
in wealth proportion $$(1 - a)$$. The standard deviation of returns, or
total risk, for this combination is:
$$\begin{aligned} \sigma ({R}_C)=[a^2 \sigma ^2({R}_{f})+(1-
a)^2 \sigma ^2 ({R}_P)+2a(1-a) Cov ({R}_{f},{R}_{P})]^{1/2}, (3.6)
\end{aligned}$$
where $$Cov ({R}_{f},{R}_{P})$$ denotes the covariance between asset
f and P returns. Since the riskless rate $$R_f$$ is constant over time, we
know that $$\sigma ^2({R}_{f}) = 0$$ and $$\rho _{fP} = 0$$ (i.e.,
the correlation of a constant with a random variable is zero). Consequently,
the term
$$Cov ({R}_{f},{R}_{P}) = \rho _{fP} \sigma ({R}_{f})\sigma
({R}_{P}) = 0$$
, where $$\rho _{fP}$$ is the correlation coefficient between their
returns. Conveniently, this result simplifies Eq. (3.6) to:
$$\begin{aligned} \sigma ({R}_C)=(1-a)^2 \sigma ^2 ({R}_P).
(3.7)
\end{aligned}$$

Fig. 3.3 Locating market portfolio as a tangent point on the ray from riskless rate

The formulas for Eqs. (3.5) and (3.6) imply that combinations of the
riskless rate and any efficient portfolio’s risky return lie on a straight line.
Figure 3.3 shows a variety of possible combinations of the riskless rate with
different risky portfolios. It is obvious that the ray from riskless rate
$$R_f$$ to the tangent point on the efficient frontier provides the
highest expected return E(R) per unit risk as measured by the standard
deviation of returns $$\sigma (R)$$. Other combinations are less
efficient than this optimal combination.
Sharpe referred to the optimal tangent point as the market
portfolio denoted M. All points on the Capital Market Line (CML) from
$$R_f$$ to M represent different proportionate investments in the
riskless asset f and market portfolio M. All risk-averse investors buy market
portfolio M, which is comprised of the value-weighted returns of all assets
in the market with weights equal to the ratio of the market value of each
asset to the aggregate market value of all assets. This theoretical result is
remarkable! What about investors that buy other risky asset portfolios?
According to Fig. 3.3, they would earn lower risk-adjusted returns than if
they invested in M.
How do investors achieve expected return and total risk combinations
along the CML? Points on the line from $$R_f$$ to M require that the
investor lend some proportion a of their funds to earn the riskless rate
 $$R_f$$ (e.g., depositing funds in an insured bank deposit account or
buying a riskless government bond) and invest the remainder in proportion
$$(1 - a)$$ in market portfolio M. Where the investor lies on this
section of the CML depends on their risk aversion as determined by the
tangency point of their indifference curve with the CML. If the investor
prefers points above M on the CML, these points could be achieved by
borrowing funds at the riskless rate $$R_f$$ and adding to their
purchases of market portfolio M.
There is a long controversy about Sharpe’s market portfolio results.
Many professional investors argue that they can “beat the market” in the
sense of actively managing investments to pick assets earning a higher rate
of return per unit risk greater than market portfolio M. Opposing this claim,
academics point to Sharpe’s mean-variance theory. Historical evidence
proves that over 90% of professional managers do not outperform common
stock market indexes over reasonable investment horizons (e.g., the S &P
500 index over five years). This empirical fact led to the widespread
adoption of index investing in well-diversified portfolios that contain all
assets in the stock market. So-called passive index funds have very low
transactions costs which help to further explain their relatively higher
performance compared to actively managed funds. Professional managers in
the latter funds seek to apply investment skills to choose portfolio assets
and incur higher operating costs from this aggressive activity (e.g.,
transactions costs, management salaries, etc.).

3.4 The CAPM

The optimal CML tangent to the market portfolio defines a simple linear
relationship between expected returns and their risk or standard deviation of
returns. How does this general equilibrium relation affect individual assets?
These assets are inefficient and lie within the investment opportunity
set somewhere below the efficient frontier. Sharpe proved that individual
assets are affected by their relationship to the market portfolio. Let’s review
his proof.5
Similar to forming combinations of the riskless asset and the market
portfolio, investors can purchase combinations of an individual risky asset i
and market portfolio M. The expected rate of return on this combination is:
(3.8)
 $$\begin{aligned} E({R}_C)=aE({R}_i)+(1-a)E({R}_M).
\end{aligned}$$
The standard deviation of returns for this portfolio is:
$$\begin{aligned} \sigma ({R}_C)=[a^2 \sigma ^2({R}_{i})+(1-
a)^2 \sigma ^2 ({R}_M)+2a(1-a) Cov ({R}_{i},{R}_{M})]^{1/2}, (3.9)
\end{aligned}$$
where $$Cov ({R}_{i},{R}_{M})$$ denotes the covariance
between asset i and M returns. Using Eq. (1.10) in Chapter 1, the covariance
term for asset i and market portfolio M can be defined as
$$Cov ({R}_{i},{R}_{M})$$ = $$\rho _{iM}$$
$$\sigma (R_i)$$ $$\sigma (R_M)$$, where
$$\rho _{iM}$$ is the correlation of returns between asset i and M,
and $$\sigma (R_i)$$ and $$\sigma (R_M)$$ are the standard
deviations of returns (or total risks) of assets i and M, respectively. Hence,
covariance is related to correlation—positive (negative) correlation will
result in positive (negative) covariance (as total risk is always positive). As
we learned earlier, the covariance term plays an important role in the total
risk of the combined portfolio C.
The slope of the efficient frontier can be derived as follows. From
Eq. (3.8), we know that $$dE({R}_C)/da =E({R}_i)-E({R}_M)$$.
Also, from Eq. (3.9), we get
$$d\sigma ({R}_C)/da = [a \sigma ^2({R}_{i})-\sigma ^2
({R}_M)+a \sigma ^2 ({R}_M)+ Cov ({R}_{i},{R}_{M})-2a\, Cov
({R}_{i},{R}_{M})]/\sigma ({R}_C)$$
. Using these results, the change in portfolio C returns per unit change in
total risk, or slope of the CML, at the market portfolio where a = 0 is:
$$\begin{aligned} \frac{dE({R}_C)}{d\sigma
({R}_C)}=\frac{[E({R}_i)-E({R}_M)]}{[-\sigma ^2 ({R}_M)+ Cov (3.10)
({R}_{i},{R}_{M})]/\sigma ({R}_M)}. \end{aligned}$$
In Fig. 3.3 we know that the slope of the CML is
$$[E({R}_M)-R_f]/\sigma ({R}_M)$$.
Setting the slopes of the CML and efficient frontier (at the market
portfolio M with $$a=0$$) equal to one another, we can write:
(3.11)
 $$\begin{aligned} \frac{[E({R}_M)-R_f]}{\sigma
({R}_M)}=\frac{[E({R}_i)-E({R}_M)]}{[-\sigma ^2 ({R}_M)+ Cov
({R}_{i},{R}_{M})]/\sigma ({R}_M)}. \end{aligned}$$
Rearranging terms yields:
$$\begin{aligned} E({R}_i)=R_f+[E({R}_M)-R_f]\frac{ Cov
(3.12)
({R}_{i},{R}_M)}{\sigma ^2({R}_M)}. \end{aligned}$$
The term $$Cov ({R}_{i},{R}_M) / \sigma ^2({R}_M)$$ plays a
pivotal role in this equation. It represents the slope of the line relating
individual asset returns to market portfolio returns. To see this more clearly,
we can substitute the parameter $$\beta _i$$ for the covariance term
to get the more commonly used simplified form of the CAPM:
$$\begin{aligned} E({R}_i)=R_f+\beta _i [E({R}_M)-R_f],
(3.13)
\end{aligned}$$
where
$$\begin{aligned} \beta _i = \frac{ Cov ({R}_{i},{R}_M)}{\sigma
(3.14)
^2({R}_M)}. \end{aligned}$$
The term $$\beta _i$$ is known as beta. It measures systematic risk,
which is the risk of the asset relative to the market portfolio. If market
portfolio M is inserted for asset i on the left-hand-side of the CAPM, its
beta would equal 1. This benchmark value can be used to determine the
relative risk of different assets. Assets with beta < 1 or beta > 1 have lower
or higher risk, respectively, than the general market as represented by the
market portfolio.
Beta risk assumes that the investor buys a diversified market portfolio
M. As such, total risk as measured by $$\sigma (R_i)$$ for the ith
assets can be broken down as follows:
$$\begin{aligned} \text {Total risk = systematic risk +
(3.15)
unsystematic risk}, \end{aligned}$$
where systematic risk is measured by beta, and unsystematic risk is an
idiosyncratic component unrelated to systematic risk. Because the
correlation of an asset with the market portfolio cannot be diversified away,
beta risk is the only risk that is priced by investors. All other idiosyncratic
risks can be eliminated by the diversifying behavior of investing in the
market portfolio containing all assets. Notice that, while idiosyncratic
risks do indeed exist, they do not affect the prices of capital assets as
determined by diversified investors. This theoretical inference complements
the earlier finding that all investors should purchase market portfolio M.

Fig. 3.4 Pricing capital assets using systematic or beta risk

Figure 3.4 gives a graphical depiction of the CAPM. The linear

relationship between the expected rate of return $$E(R_i)$$ and beta
risk $$\beta _i$$ for individual assets is referred to as the Security
Market Line. The market portfolio M has expected rate of return
$$E(R_M)$$ and systematic risk $$\beta = 1$$. Notice that the
market risk premium or return in excess of the riskless rate equals
$$E(R_M) - R_f$$. From Eq. (3.13), the risk premium for an
individual asset, or $$E(R_i) - R_f$$, equals beta times the market
risk premium, or $$\beta _i [E({R}_M)-R_f]$$. These basic
relationships describe the CAPM.

Fig. 3.5 Expected rates of return and beta risk

As an example, Fig. 3.5 illustrates how expected returns change with

beta risk. We assume that the riskless rate $$R_f$$ = 4% and
expected market portfolio rate of return $$E(R_M) = 9\%$$, such that
the market risk premium $$E(R_M) - R_f = 5\%$$. The table shows
the expected rate of return for asset i, or $$E(R_i)$$, for different beta
values, or $$\beta _i$$. Again, the beta of the market portfolio at one
is a benchmark against which to evaluate the risk of different assets. Assets
with betas less (more) than one, have less (more) risk than the market and
commensurately lower (higher) expected returns. All $$E(R_i)$$
increase linearly with beta and lie on the SML. Assets in disequilibrium
located above (below) the SML are underpriced (overpriced). Arbiteurs will
buy (sell) these mispriced assets until they lie on the equilibrium SML.
We can write the CAPM in alternative forms that yield insights into the
valuation of assets. We begin by defining the risky price of the ith asset at
the end of a single-period as $$P_{i1}$$, where we now use a time
subscript t. This risky price is a random variable. The asset’s current price
$$P_{i0}$$ is known and so is a constant. Referring back to Eq.
(3.16) for the ex post return, we can similarly define the ex ante rate of
return for the ith asset as:
$$\begin{aligned} E(R_{i}) =\frac{E(P_{i1}) - P_{i0}}{P_{i0}},
(3.16)
\end{aligned}$$
where $$E(P_{i1})$$ is the expected future price of the asset at the
end of the period at time 1, and $$P_{i0}$$ is the current equilibrium
price of the asset at time 0. Here the time subscripts $$t = 1$$ and
$$t = 0$$ only are used for prices, and no subscript is again used for
single-period returns. Substituting this one-period return into the left-hand-
side of CAPM relation (3.13), we have:
$$\begin{aligned} \frac{E(P_{i1}) - P_{i0}}
{P_{i0}}=\frac{E(P_{i1})}{P_{i0}} - 1 = R_f+\beta _i (3.17)
[E({R}_{M})-R_f]. \end{aligned}$$
Using some algebra, we can re-write this equation as:
$$\begin{aligned} P_{i0}=\frac{E(P_{i1})}{1+R_f+\beta _i
(3.18)
[E({R}_M)-R_f]}. \end{aligned}$$
Equation (3.18) is known as the risk-adjusted rate of return valuation
formula.6 The risky future price or cash flow $$E(P_{i1})$$ in the
numerator is divided by the risky discount rate
$$1+R_f+\beta _i [E({R}_M)-R_f]$$ in the denominator, where risk
premium $$\beta _i [E({R}_M)-R_f]$$ is added to the riskless one-
period discount rate $$1 + R_f$$. If the asset has no risk, then this
valuation equation reduces to
$$P_{i0} = E(P_{i1})/(1+R_f) = P_{i1} /(1 + R_f)$$, as
$$E(P_{i1}) = P_{i1}$$ when $$P_{i1}$$ is riskless.
Alternatively, if we define
$$\beta _i = Cov (R_i, R_M) / \sigma ^2(R_M)$$ as before, and
substitute $$[E(P_{i1}) - P_{i0}]/P_{i0}$$ for $$R_i$$, beta
can be re-written as:
(3.19)
 $$\begin{aligned} \beta _i = \frac{1/P_{i0} Cov [(P_{i1} -
P_{i0}), R_M]}{\sigma ^2(R_M)}. \end{aligned}$$
Because $$P_{i1}$$ is a random risky price, and $$P_{i0}$$ is
the current price (a constant) with zero correlation with $$R_M$$,
beta reduces to
$$\begin{aligned} \beta _i = \frac{1/P_{i0} Cov (P_{i1}, R_M)}
(3.20)
{\sigma ^2(R_M)} \end{aligned}$$
or $$\beta _i = Cov (P_{i1}, R_M) / P_{i0} \sigma ^2(R_M)$$. Next,
we replace beta with this value in Eq. (3.18) to get
$$\begin{aligned} P_{i0}=\frac{E(P_{i1})}{1+R_f + [ Cov
(P_{i1}, R_M) / P_0 \sigma ^2(R_M)] [E(R_M)-R_f]}. (3.21)
\end{aligned}$$
Again rearranging terms, we finally obtain the following certainty
equivalent valuation formula:
$$\begin{aligned} P_{i0}=\frac{E(P_{i1}) - [ Cov (P_{i1},
(3.22)
R_M)/\sigma ^2(R_M)][E(R_M)-R_f]}{1 + R_f}. \end{aligned}$$
Here the certainty equivalent cash flow in the numerator is discounted at the
riskless rate to yield the equilibrium current price of the asset
$$P_{i0}$$. Unlike the risk-adjusted rate of return valuation formula
(3.18), wherein the risky discount rate in the denominator is adjusted
upward by adding the risk premium to the riskless rate, cash flows in the
numerator are adjusted for risk by subtracting the risk premium from the
risky future price. In both of these alternative CAPM models, the value of
an asset is determined by its expected future cash flow and beta risk as well
as the market risk premium and riskless rate. Given homogeneous
expectations among investors about these determinants, no individual risk
preferences (as reflected by utility functions) affect asset value.

3.5 Implications of the CAPM

The CAPM is an elegantly simple theoretical model based on previously
established portfolio theory and risk preference theories by Markowitz and
Tobin. As mentioned earlier, Treynor, Lintner, and Mossin developed
comparable versions of the CAPM. Unlike Sharpe, they allowed short sales
of assets, which helps to make the model more general and applicable to the
real world. Also, unlike Sharpe who allowed more than one risky portfolio
to be efficient, the other authors found that tangent portfolio M is the only
efficient risky portfolio—that is, M is located on the highest return/risk
CML from the riskless rate that represents the efficient frontier in the
presence of a riskless asset. In his paper, Sharpe argued that one asset might
be in one efficient risky portfolio and a second asset could be in another
efficient risky portfolio. Because all efficient risky portfolios are perfectly
correlated, he inferred that the betas of all assets would be unchanged with
respect to alternative efficient portfolios. However, Fama (1968) clarified
this issue by observing that market portfolio M must contain all assets.
Since any assets not in M would not be held by investors, this possibility is
inconsistent with equilibrium in which markets are clear for all assets. In
the CAPM, market portfolio M contains all assets in proportion to their
market values relative to total or aggregate market value.
Treynor (1961, 1962) never published his papers but over time received
credit for being the first person to derive the CAPM.7 He had given a copy
of his first paper to Lintner at Harvard in 1960. As a matter of record, he
publicly presented his work in 1962 to MIT university faculty in a research
seminar. You can find his papers on the internet.8
Lintner (1965) considered different idealistic assumptions underlying
the CAPM and concluded that the CAPM would hold under more complex
real-world conditions. For example, he argued that the CAPM would hold
even if borrowing rates exceeded lending rates. Securities firms typically
charge clients a rate of interest greater than the riskless government interest
rate on borrowed funds to invest in securities. However, this objection
implies only that there is a kinked CML with a higher slope from
$$R_f$$ to M and a lower slope (due to borrowing rates being greater
than lending rates) above M. And, some large investors may well be able to
borrow at close to the riskless government rate (e.g., global banks,
insurance, securities firms, etc.). Hence, Lintner inferred that the CAPM,
while theoretically simplistic, was applicable to real-world decision-
making, including capital budgeting and other financial problems.9
Mossin (1966) cited Sharpe’s CML in his analysis. He employed the
CML to define the “price of risk” for individual assets. This price was
defined as the compensation for risk per unit of an individual asset’s risk. In
effect, similar to Sharpe’s SML, he derived the asset pricing relation for
individual assets from the general market equilibrium CML.
The significance of the CAPM rests in its implications to the theory and
practice of finance. The following provides a short list of major
implications.
1.
All assets lie somewhere along the SML. Assets with prices above
(below) the SML will be bought (sold) by arbiteurs until their prices
change and fall back to the SML. Hence, the SML represents an
equilibrium relationship between risk and return in the marketplace. It
can be used to determine whether assets are over- or underpriced.
2.
Market beta can be used to rank order stocks or portfolios by their risk.
Returns for assets should be related to their beta risk levels. Investors
can expect lower (higher) returns on assets with lower (higher) beta
risk in portfolio selection decisions. For portfolios, assuming the
investor has $$x\%$$ of their wealth in asset X and $$y\%$$ in
asset Y, the beta of the portfolio will be the simple weighted average of
their betas, or $$\beta _p = x\beta _X + y\beta _Y$$.10
3.
Investors do not need to spend countless hours identifying and
measuring credit risk, liquidity risk, legal risk, management risk, etc.,
of individual stocks. All of these risks are incorporated in beta risk,
which summarizes the relevant risk facing diversified investors holding
the market portfolio.
4.
Given beta for a stock, managers of firms can estimate the cost of
equity and cost of capital (i.e., the cost of capital is the weighted costs
of debt and equity capital). In turn, these costs can be used to select the
highest net present value (NPV) projects in capital budgeting decisions.
5.
Idiosyncratic risks unrelated to systematic risk will earn only the
riskless rate of return. Only assets with positive systematic risk will
have expected rates of return greater than the riskless rate.
6. Market imperfections, including perfect capital market frictions such as
taxes, transactions costs, institutional structure, government regulation,
etc., do not significantly affect the main results of the CAPM.
However, real-world empirical tests are needed to confirm this
 However, real world empirical tests are needed to confirm this
implication.
7.
Beta measures the sensitivity of an asset’s rate of return to changes in
the market rate of return. Since the market return is affected by general
economic activity, asset returns are sensitive to fundamental economic
forces also.

8.
Investors are able to purchase a portfolio in the marketplace that is a
close substitute or proxy for market portfolio M.
9.
There are no other systematic risk factors than the market factor equal
to the market risk premium. No mispricing exists after taking into
account sensitivity to the market factor as measured by beta risk. Tests
of other potential risk factors are needed to support this implication.
As we will see in forthcoming chapters, these implications motivated a
tremendous amount of research in the years to come. Can such a simple
model really be valid? Will it work in the real world? Does the CAPM lead
to other theoretical discoveries? Are there competing models of asset
pricing? As we will see, confirming the final remarks of Lintner at the end
of his paper, the CAPM opened the doors to a flood of new theories and
applied for empirical work.

3.6 Summary
The traditional approach to valuation employs the present value formula.
This formula specifies asset prices as a function of expected future cash
flows divided by expected discount rates. Fundamental accounting, finance,
and operating information about the firm are used to estimate expected cash
flows. A wide variety of risks are incorporated into the expected discount
rate. Alternatively, historical rates of return on an asset can be used to
approximate expected discount rates. However, historical rates have the
drawback of being backward looking and, therefore, are not helpful in
determining whether an asset is over- or undervalued relative to future
returns.
The CAPM represents a breakthrough in terms of measuring the
relevant risk to diversified investors. Based on idealistic assumptions of
perfect competition, homogeneous expectations, utility maximization, risk
aversion, and riskless borrowing and lending rates, Treynor, Sharpe,
Lintner, and Mossin independently derived the price of risky assets in
equilibrium. To do this, they assumed that Markowitz portfolio theory and
the Tobin Separation Principle hold. Investors that practice Markowitz
diversification will hold a single optimal portfolio on the efficient frontier.
This efficient portfolio is obtained as the tangent point on a ray extending
from the riskless rate to the efficient frontier. The resultant ray is the Capital
Market Line (CML) relating expected rates of return to total risk as
measured by the standard deviation of returns. All investors will purchase
the tangent point known as the market portfolio M, which is a separate
decision from their risk preferences. Depending on their risk preferences,
investors will hold different combinations of the riskless rate and the market
portfolio. Investors who want to take more risk than the market portfolio
can borrow funds and increase their holdings of the market portfolio. These
results are logical outcomes of Markowitz diversification and Tobin risk
preference concepts.
The next step is to derive the price of an individual asset. By setting the
slopes of the linear CML to the efficient frontier of risky assets, Sharpe
showed that asset prices are determined by the riskless rate, expected rate of
return on the market portfolio, and beta risk. Beta risk (denoted
$$\beta$$) is a parameter relating the expected rate of return of asset i
to the market risk premium defined as the expected rate of return on the
market portfolio return minus the riskless rate. The linear relationship
between expected rates of return for assets and their beta risks is the
Security Market Line (SML). The SML represents the equilibrium pricing
relationship in which only beta measuring systematic risk is priced by
investors. Hence, total risk as measured by the standard deviation of
returns can be broken down into undiversifiable systematic risk (beta) and
diversifiable unsystematic (idiosyncratic) risk. Assets that are underpriced
(overpriced) lie above (below) the SML; arbiteurs will buy (sell) these
assets until they lie on the SML. Thus, the SML is an equilibrium
return/risk relation. Two alternative but equivalent forms of the CAPM are
the risk-adjusted rate of return valuation formula and certainty equivalent
valuation formula. Both of these formulas, like the CAPM itself, value
assets without regard to the risk preferences of individuals that have
homogeneous expectations. If investors have heterogeneous expectations, a
single efficient market portfolio may not be identified by investors, which is
critical to the CAPM. For this reason, empirical tests of the CAPM
discussed in the next chapter jointly test the CAPM and the efficiency of the
market portfolio.
The CAPM has many important implications. All investors buy the
same market portfolio. All systematic risk is captured in beta risk related to
the market portfolio. Besides identifying over- and underpriced assets, it
can be used in portfolio management, capital budgeting, and other financial
decisions. Its assumptions are idealistic but may well be relaxed in many
cases, such as the allowance of risky borrowing rates and short sales. The
simple elegance of the CAPM is both an advantage and disadvantage. If
true, it allows investors to spend less time evaluating a myriad of different
risks affecting assets (e.g., credit risk, liquidity risk, legal risks, regulation,
etc.) and instead focus on beta risk. Contrary to the CAPM, many
professional investment managers believe that they can outperform the
market portfolio (e.g., a general market index) by means of skill in
evaluating market information. As will see in forthcoming chapters, the
CAPM opened the door to a large body of theoretical and applied research.
Even today, researchers grapple with what we might call the CAPM
paradox—are asset prices determined by the equilibrium mechanics of
Treynor–Sharpe–Lintner–Mossin, or are they entangled in a more complex
pricing mechanism?

Questions
1.
Write the present value formula for future cash flows. Re-write this
present value formula using m-talk. Redefine the m-talk equation in
terms of expected returns.
2.
Draw the Capital Market Line (CML) of Sharpe (1964). What does
this line show?
3.
Assuming Markowitz’s efficient frontier and Tobin’s mean-variance
indifference curves, what portfolio should the investor select?
Illustrate this portfolio with a graph with mean returns on the Y-axis
and standard deviation of returns on the X-axis.
4. Write the expected return of the portfolio combination of the riskless
 asset in proportion a and risky efficient portfolio P in proportion (
$$1-a$$). What is the standard deviation of returns for this
combination? Since the riskless rate $$R_f$$ is a constant over
time, to what does this standard deviation reduce? What do these
results imply to investors?
5.
Given the results for question 4 above, which portfolio will an
investor optimally select to maximize their expected returns per unit
risk as measured by the standard deviation of returns
$$\sigma (R)$$? Draw a picture of your Capital Market
Line (CML) solution. How can investors locate themselves at
different points along this CML?
6.
Do you think that professional investors can buy a risky portfolio that
“beats the market” in the sense of outperforming portfolio M?
7.
Draw the Security Market Line (SML). Is this line an equilibrium
relation between risk and return? Is systematic risk part of total risk?
8.
A certainty equivalent form of the CAPM is as follows:
$$\begin{aligned} P_0=\frac{E(P_1) - [ Cov (P_1, R_M)/\sigma
^2(R_M)][E(R_M)-R_f]}{1 + R_f}.\nonumber \end{aligned}$$
Why does the equation use the riskless rate to discount the cash flows
in the numerator?
9.
Give a definition of the market portfolio M. Are some assets in the
market not contained in M?
10.
The CAPM has a number of implications for the theory and practice
of finance. Discuss five implications.

Problems
1. Assume that investors hold a portfolio of an individual risky asset and
riskless asset f. Starting with the expected return of this combination
and its standard deviation, derive the slope of the efficient frontier and
then set this slope equal to the slope of the efficient frontier at market
 portfolio M. Rearranging terms, write the equation for the CAPM in
more general form as well its more common simplified form.
2.
Using data in the following table, compute the expected returns for
different assets using the CAPM.

$$E(R_i)$$ $$R_f$$ (%) Beta, $$\beta _i$$ $$E(R_M)-R_f$$ (%)

2.00 -0.25 4.00
2.00 0.00 4.00
2.00 0.25 4.00
2.00 0.75 4.00
2.00 1.00 4.00
2.00 1.50 4.00
2.00 2.00 4.00

References
Black, F. 1981. An open letter to Jack Treynor. Financial Analysts Journal 37: 14.

Cochrane, J. H. 2005. Asset Pricing: Revised Edition. NJ: Princeton University Press, Princeton.

Fama, E.F. 1968. Risk, return, and equilibrium: Some clarifying comments. Journal of Finance 23:
29–40.
[Crossref]

Ferson, W. E. 2019. Empirical Asset Pricing: Models and Methods. MA: The MIT Press, Cambridge.

French, C.W. 2003. The Treynor Capital Asset Pricing Model. Journal of Investment Management 1:
60–72.

Lintner, J. 1965. The valuation of risk assets and the selection of risky investments in stock portfolios
and capital budgets. Review of Economics and Statistics 47: 13–37.
[Crossref]

Markowitz, H. 1952. Portfolio selection. Journal of Finance 7: 77–91.

Markowitz, H.M. 1959. Portfolio Selection: Efficient Diversification of Investments. New York, NY:
John Wiley & Sons.

Mossin, J. 1966. Equilibrium in a capital asset market. Econometrica 34: 768–783.

[Crossref]

Sharpe, W.F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance 19: 425–442.
Tobin, J. 1958. Liquidity preference as behavior toward risk. Review of Economic Studies 4: 65–86.
[Crossref]

Treynor, J.L. 1961. Market value, time, and risk. Unpublished manuscript.

Treynor, J.L. 1962. Toward a theory of market value of risky assets. Unpublished manuscript.

Weston, T.E., and J.F. Weston. 1980. Financial Theory and Corporate Policy. Reading, MA:
Addison-Wesley Publishing Company.

Footnotes
1 In different contexts, the variable m is sometimes referred to as the asset pricing kernel or
marginal rate of substitution.

2 See advanced textbooks by Cochrane (2005) and Ferson (2019) for excellent in-depth discussions
and applications of m-talk in asset pricing models.

3 For example, assuming no risk, if you indifferent between $$\$100$$ today or $$\$101$$ at
the end of a year, your rate of time preference is 1%. This rate can be attributed to your
utility preference of present consumption over future consumption.

4 See Sharpe (1964, p. 427 footnote 7).

5 See Sharpe (1964, p. 438, footnote 22).

6 See Copeland and Weston (1980, p. 170).

7 See Black (1981).

8 See the following Social Science Network (SSRN) websites:

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2600356 and
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2600356.
9 Linter also did not assume that all investors agree on the distribution of asset returns in the next
period; instead, he argued that asset prices are based on the weighted average of investor expectations
about asset returns.

10 Note that, to find Markowitz efficient portfolios, it is not necessary to use optimization methods
such as quadratic programs. Conceptually at least, beta could be used to construct efficient
portfolios with different levels of market risk.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_4

4. The Market Model

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_4

The famed capital asset pricing model (CAPM) of Treynor (1961; 1962),
Sharpe (1964), Lintner (1965), and Mossin (1966) is surprisingly elegant in
design. It hypothesizes that only a single market factor is sufficient to
measure systematic risk and price assets in financial markets. All investors
will price assets using the market factor regardless of their individual risk
preferences. To test this hypothesis, the market model was invented in the
form of a simple regression equation. Using a proxy for the market portfolio
and a riskless rate, this regression model enables the estimation of beta
risk for assets.
The earliest applications of the market model focused on the U.S. stock
market. As we will see, this early evidence did not entirely support the
CAPM. High (low) beta stocks had lower (higher) returns than predicted by
the CAPM. In effect, the slope of the Security Market Line (SML) was
flatter than predicted by the theory. This flatter SML resulted in a higher
than expected intercept term, which implies that other risks may well exist
that are embedded in the intercept or $$\alpha$$ parameter in the
market model. Given that $$\alpha$$ is fairly large in magnitude, as
discussed in forthcoming chapters, researchers subsequently proposed
various remedies that led to new forms of the CAPM as well as alternative
models extending the CAPM to multiple factors.
In this chapter, we take a look at the market model and early CAPM
evidence. Because the market model is a statistical model, we supplement
the discussion with a primer on simple regression analysis, including the
interpretation of regression coefficients, justification for ordinary least
squares (OLS) estimation of regression coefficients, relation of the market
model to the theoretical CAPM, and goodness-of-fit of regression models.

4.1 Empirical Form of the CAPM

From Chapter 3, the theoretical CAPM specifies the expected rate of
return on the ith asset in ex ante form as follows:
$$\begin{aligned} E(R_i)=R_f+\beta _i [E(R_M)-R_f].
(4.1)
\end{aligned}$$
where $$R_f$$ is the riskless rate, $$E(R_M)$$ is the expected rate of
return on the market portfolio, and $$\beta _i$$ measures systematic beta
risk. How can we estimate this model in the real world?
To investigate the validity of the CAPM, an empirical model is needed.
For this purpose, the variables in the CAPM need to be proxied with data
available in financial markets. It is reasonable to proxy the riskless rate by
using a high-quality government bond rate that has little or no default risk.
The rate of return on the market portfolio M is a more difficult variable to
proxy. Since M theoretically contains all assets by their market values,
when applying the model to the stock market, researchers proposed using a
general value-weighted market index. For example, the S &P 500 index
contains the largest 500 stocks that comprise a large proportion of the total
market capitalization of the U.S. stock market. A more comprehensive
aggregate stock market portfolio is the CRSP index containing all U.S.
stocks that are provided by the Center for Research and Security Prices at
the University of Chicago.
Markowitz (1959) originally proposed an empirical model for the
CAPM that Sharpe (1963) and Fama (1968) further developed. They
proposed the following so-called market model:1
$$\begin{aligned} R_{it}-R_{ft}=\alpha _i+\beta _i (R_{mt}-
(4.2)
R_{ft}), \end{aligned}$$
where $$R_{it} - R_{ft}$$ is the excess realized rate of return on
asset i in time period t (e.g., daily or monthly), and
$$R_{mt} - R_{ft}$$ is the realized excess average rate of return on
the stock market as a whole. Notice that an intercept term, or
$$\alpha$$ parameter, is included to denote the excess realized rate of
return assuming no beta risk (i.e., $$\beta = 0$$). According to
CAPM theory, $$\alpha$$ should be zero. Also, we use the lower case
notation m for the stock market portfolio, which is a proxy for the
theoretical (unobservable) market portfolio M.
The market model is a simple regression model of y (excess asset
returns) on x (excess market returns). Typically, ordinary least squares
(OLS) regression methods are used for estimation purposes. OLS is the
most widely used type of regression analysis that is available in Excel and
all statistical packages (e.g., Stata, Matlab, etc.). Free software on the
internet is available for R and Python program codes to estimate OLS
regression models. Forthcoming Sect. 4.4 provides a primer on regression
analysis.
Jensen (1968) recognized that $$\alpha _i$$ enables a test of the
CAPM. So-called Jensen’s alpha is a joint test of both the efficiency of the
proxy for the market portfolio and the possibility of mispricing due to
missing factors. If the empirical market portfolio m is not a good proxy for
the theoretical market portfolio M, then $$\alpha$$ would be nonzero.
Roll (1977) argued that the CAPM cannot be tested without an efficient
market proxy for M, which is known as the Roll critique. Moreover, if one
or more other factors are important in pricing stocks beyond the market
excess return factor, again $$\alpha$$ would not equal zero. Given the
simplicity of the CAPM, the possibility of missing factors is certainly
plausible.
 In the real world, all stocks will not lie exactly on the Security Market
Line (SML) relating asset returns to beta risks. Individual stocks’ returns
will randomly be located above or below the SML due to random
movements in their stock prices (rather than over- or underpricing). Taking
into account random stock prices, the market model can be modified as
follows:
$$\begin{aligned} R_{it}-R_f=\alpha _i+\beta _i (R_{mt}-
(4.3)
R_{ft})+e_{it}, \end{aligned}$$
where $$e_{it}$$ is a random error term (with mean zero and
constant variance) that captures the scatter of assets’ returns around the
SML.

4.2 Early Tests of the Market Model

In this section we review some of the early evidence on the market model.
Overall, stock market data suggested that the Security Market Line
(SML) relating stock returns to beta risk was flatter than CAPM theory
would predict. The slope of the SML was lower and the intercept was
higher than hypothesized. This evidence caused considerable controversy
about the CAPM. Were the data in the tests sufficient? Were the statistical
tests appropriate? Was the CAPM too simplistic? These and other questions
were addressed by researchers in the early studies.

4.2.1 Black, Jensen, and Scholes Study

Black, Jensen, and Scholes (BJS) (1972) proposed a direct test of the
CAPM using the market model. According to the CAPM, the $$\alpha$$
intercept should be zero. Their data consisted of monthly returns for U.S.
stocks trading on the New York Stock Exchange (NYSE) and U.S.
government interest rates (i.e., 30-day Treasury rates) for the period 1926 to
1966.2 The study included between 600 to 1,000 individual stocks in
different years.
BJS is a landmark study that helped to establish the use of portfolios in
asset pricing tests. They discovered that individual stocks should not be
used in CAPM tests due to substantial measurement errors in estimating
betas and, in turn, intercepts. These measurement errors tend to cancel one
another out in portfolios. Consequently, their portfolio sorting method
became the standard approach in later asset pricing tests.3 One drawback of
forming portfolios is that the range of beta values is decreased compared to
their range among individual stocks. This more narrow range of betas
makes it more difficult to detect a significant relation between stock returns
and beta. In the extreme, if random portfolios were used, the portfolios
would tend to have mean returns similar to that of the market portfolio
proxy. In this case, all the alphas ( $$\alpha$$s) would be zero with
betas ( $$\beta$$s) equal to approximately one. Of course, this test
would be meaningless; no information on how beta affects returns would be
gained. By sorting stocks into portfolios by their betas, BJS mitigated this
potential pitfall but nonetheless constrained the full range of return/beta
relations in the population of stocks.
They grouped stocks into 10 portfolios sorted on betas. In forming beta
groups of stocks, they employed five years of monthly returns to estimate
beta for all individual stocks. For example, the five-year period 1926 to
1930 was used to estimate individual stocks’ betas. After forming 10 beta
groups, their monthly returns in the next (out-of-sample) 12 months of 1931
were computed and their beta group was retained. They then rolled forward
this process by one year, estimated betas for the five-year period 1927 to
1931, formed 10 beta groups, and computed the monthly returns in the next
year 1932 for each beta group. This process was repeated to construct a 35-
year time series of monthly returns from 1931 to 1965 (i.e., 420
observations). These time-series returns were used to estimate the market
model Eq. (4.3) for each of the 10 beta portfolios. On average, the betas of
the portfolios ranged from about 0.5 to 1.5 with some variation in this range
from year-to-year in the sample period.
For the entire sample period, the results of BJS market model tests
followed a definite pattern. Stock portfolios with betas greater (less) than
one had negative (positive) intercepts. To evaluate the CAPM, a t-test of the
intercept for each of the 10 portfolios was implemented as follows:
$$\begin{aligned} t = \frac{\hat{\alpha }_p-0}{s_{\hat{\alpha
(4.4)
}_p}}, \end{aligned}$$
where $$\hat{\alpha }_p$$ is the OLS regression estimate for each
portfolio p, $$s_{\hat{\alpha }_p}$$ is the OLS standard error of
$$\hat{\alpha }_p$$, and 0 is the population value of
$$\alpha$$ according to CAPM theory. Note that
 $$\hat{\alpha }_p$$ and $$s_{\hat{\alpha }_p}$$ are estimated
for each portfolio using an OLS regression of the market model with
monthly returns from 1931 to 1965. The t-test above assumes that the error
terms of the market model are normally distributed.4 The results of these t-
tests for the 10 portfolios showed that 3-out-of-10 tests were significant at
the 10% level. Stock portfolios with $$\beta$$s nearer to one had
insignificant $$\alpha$$s. But portfolios with very high or low
$$\beta$$s had significant $$\alpha$$s. Again, CAPM theory
hypothesizes that no significant $$\alpha$$s should occur; hence,
some evidence was contrary to the CAPM. While a definite pattern in
positive and negative $$\alpha$$s was obvious, they cautiously
inferred that t-tests for the full sample of all stocks were not sufficient to
make a clear interpretation of the results.
To more fully investigate their sample data and increase the number of
tests, BJS divided their sample period into four subperiods with 105 months
each. In the first period, the results flipped: stock portfolios with
$$\beta$$s greater (less) than one had positive (negative) intercepts.
In the other three subperiods, the results flipped back to agree with the total
sample results: stock portfolios with betas greater (less) than one had
negative (positive) intercepts. In the last two subperiods, the t-tests were
stronger, with both higher negative and positive t-statistics associated with
$$\beta$$ portfolios having high and low $$\alpha$$s,
respectively. Related to the latter findings for the last two subperiods,
average $$\beta$$ estimates for the 10 portfolios suggested that the
slope of the SML was flatter than in the other subperiods. Thus, low
$$\beta$$ stocks had higher returns than predicted, and high
$$\beta$$ stocks had lower returns than predicted. No statistical tests
of the $$\beta$$s were conducted using time-series monthly returns in
the market model. To test the $$\beta$$s, they conducted cross-
sectional regression tests that will be discussed shortly with respect to the
Fama and MacBeth (1973) study.
It is noteworthy that the magnitudes of average $$\alpha$$s for
portfolios were relatively high at between about $$-0.50$$% and
$$+0.35$$% per month. In other words, on an annual basis, beta
risk in the market model did not explain between $$-6$$% and
 $$+4$$% of stock returns in different beta portfolios. For comparison,
the average market return on all NYSE stocks was 1.42% per month, or
about 17% per year in their sample period. Hence, the magnitudes of
average $$\alpha$$s suggested that something more than beta was
needed to complete price stocks.
Based on their findings, BJS inferred the following:
1.
the SML appears to be flatter (or less upward sloping) for the beta-
sorted portfolios than predicted by CAPM theory;
2.
portfolios with betas greater (less) than one tended to have negative
(positive) $$\alpha _p$$ intercepts; and
3.
there is some instability or nonstationarity of $$\alpha$$ and
$$\beta$$ portfolio estimates that determines the shape of the SML
over time.
Taken together, they concluded that the results rejected the CAPM. Due
to this failure, in an effort to account for the large magnitudes of the
$$\alpha$$ estimates discussed above, further tests were conducted by
BJS on a second missing factor based on the theoretical CAPM model of
Black (1972). In the next chapter, we cover Black’s more general CAPM
that contains a second factor uncorrelated with the market factor and
discuss the BJS findings.

4.2.2 Fama and MacBeth Study

A number of studies confirmed the BJS finding that the SML was flatter
than predicted by the CAPM, including Douglas (1969), Friend and Blume
(1970), Miller and Scholes (1972), Fama and MacBeth (1973),
Stambaugh (1982), Gibbons and Shanken (1983), and others. Here we take
a closer look at the Fama and MacBeth (FM) study , which became famous
in asset pricing tests due to its methodological advances.
FM gathered monthly returns for NYSE stocks in a longer sample
period from January 1926 to June 1963. Like BJS, they used portfolios in
their tests of the market model due to the potential error-in-the-
variables problem that exists for individual stock estimates of beta. Since
the market portfolio proxy m is not the same as the latent (or unobserved)
true market portfolio M, beta is estimated with error. If these errors in beta
estimates are not perfectly positively correlated, they will tend to cancel out
in portfolios, which provides a more accurate measurement of beta.
Consequently, they formed 20 beta-sorted portfolios and constructed their
equal-weighted average returns over time.
To test the CAPM, FM set up the following novel cross-sectional
regression model:
$$\begin{aligned} R_{pt}=\alpha _0+\lambda _{m} \beta _{pt-
1}+\lambda _{nl} \beta ^2_{pt-1}+\lambda _s \bar{\sigma }(e)_{pt-1} (4.5)
+e_{pt}, \end{aligned}$$
where $$R_{pt}$$ is the portfolio return in month t,
$$\alpha _0$$ is the intercept term not explained by the other
variables in the equation, $$\beta _{pt-1}$$ is the portfolio beta
estimated using monthly returns in previous years (e.g., four or five years),
$$\beta ^2_{pt-1}$$ is used to test for a possible non-linear
relation between returns and beta risk, $$\bar{\sigma }(e)_{pt-1}$$ is
the average of the standard deviation of monthly residual errors for stocks
in the portfolio in previous years, and $$e_{pt}$$ are independent
and normally distributed error terms.5 Independent variables on the right-
hand-side of the regression model are measured prior to the portfolio
returns on the left-hand-side. This out-of-sample approach to testing the
CAPM is predictive. FM argued that normative models, such as
Markowitz’s (1959) portfolio theory that was intended to help investors
make better decisions, are only valid to the extent that past information is
related to future returns.
Importantly, the $$\lambda$$ terms in Eq. (4.5) are known as
market prices of risk associated with the $$\beta$$ estimates (often
called loadings). The cross-sectional regression coefficient
$$\lambda _m$$ estimates the average market price of risk (or market
risk premium) equal to $$\bar{R}_{mt}-\bar{R}_{ft}$$ in the
sample period. We can compare this estimate to the actual average market
risk premium in the sample period. Of course, if the CAPM holds, the
estimated and actual market prices of beta risk should be approximately
equal to one another. If non-linearity exists, then the market price of
squared $$\beta$$, or $$\lambda _{nl}$$, will be nonzero. And,
if market price $$\lambda _s$$ is nonzero, some other non-beta risk
must exist that is not captured by beta.
More generally, we can summarize the above two-step testing approach
as follows:
1.
In the first step, time-series regressions are estimated to obtain
$$\beta$$ coefficients or other risk loadings in an estimation
window (e.g., five years of monthly returns or one year of daily
returns).
2.
In the second step, cross-sectional regressions are estimated based on
regressing step ahead returns (e.g., one-month-ahead) on $$\beta$$s
from the first step, which provides coefficient estimates of the
$$\lambda$$ risk premiums associated with $$\beta$$ and other
loadings.
Most studies using this two-step approach roll forward one month at a
time and repeat these steps over and over until the end of the sample period.
A time series of $$\lambda$$ risk premium estimates (i.e., the
coefficients in second step cross-sectional regressions) is generated for each
$$\beta$$ or other risk parameter. Using these monthly risk premium
estimates, the average $$\lambda$$ is computed and then a t-test of
this average risk premium is performed. The t-test seeks to determine if the
average $$\lambda$$ risk premium is significantly different from
zero. 6

Let’s review FM’s empirical results. They ran different versions of

regression Eq. (4.5) that use all variables in the model as well as subsets of
these variables (e.g., drop the squared $$\beta$$ variable). First, no
evidence of non-linearity of $$\beta$$ was found, as the t-tests of
$$\lambda _{nl}$$ were not significantly different from zero. They
inferred that the relation between expected returns and beta risk is linear.
Also, t-tests were insignificant for $$\lambda _s$$, which suggests
that non-beta risks are not present. In their test period 1935 to 1968, the
market price of beta risk, or $$\lambda _m$$, had a highly significant
t-statistic. In subperiods, this t-statistic was smaller and not significant in a
number of instances. Estimates of $$\lambda _m$$ were positive
which is consistent with the CAPM relation between returns and beta risk.
FM concluded that their findings supported the notion that beta risk is
priced by investors in the capital market. As discussed earlier, this finding
also implies that the market portfolio proxy m is efficient.
Lastly, FM tested whether the market price of risk equaled the market
portfolio proxy return minus the riskless government interest rate. Using
their cross-sectional test, they compared the market price of beta risk, or
$$\lambda _m$$, to actual $$\bar{R}_{mt}-\bar{R}_{ft}$$ in
different periods. In some subperiods these two estimates were similar to
one another, but in other subperiods they found that
$$\lambda _m &lt; \bar{R}_{mt}-\bar{R}_{ft}$$. The latter result
means that a rate of return greater than the riskless rate should be used in
the market risk premium. In other words, the market risk premium should
be defined as $$E(R_{mt}) - E(R_{0t})$$, where
$$E(R_{0t}) &gt; R_{ft}$$. Here the benchmark rate of return
$$E(R_{0t})$$ contains some risk premium in excess of the riskless
rate.
Further tests were conducted using time-series regressions of portfolio
returns regressed on excess market returns, such that the intercept
term should equal the riskless government rate. These tests confirmed that
the intercept term was greater than the government rate. Importantly, FM’s
results help to explain the BJS findings discussed earlier. The slope of the
SML is flatter and the intercept is higher than predicted by the CAPM due
to this higher benchmark rate that should be used in estimating market risk
premiums. Also, this result set the stage for the next innovation in the
development of the CAPM. As presented in upcoming Chapter 5, Black
(1972) proposed that $$E(R_{0t})$$ is the expected rate of return on
a zero-beta portfolio that is greater than the riskless rate $$R_{ft}$$
and (like the riskless rate) uncorrelated with the expected market portfolio
return $$E(R_{Mt})$$.

4.3 Estimating the Market Model

In this section we collect some real-world stock return data and estimate the
market model. Monthly stock returns for Microsoft (ticker symbol MSFT)
are analyzed. The market portfolio proxy is the Standard & Poor’s (S &P)
500 index containing the largest 500 stocks in the U.S. stock market by
market capitalization (i.e., number of shares outstanding times the current
market price). This market index typically accounts for approximately 70–
80% of the total market capitalization of all stocks trading in the U.S.
market. We use adjusted closing prices that take into account dividends and
stock splits. Monthly Treasury bill rates for one-month Tbills are
downloaded from the FRED database of the Federal Reserve Bank of St.
Louis available online.7 We gather returns for the 36-month period from
January 2018 to December 2020. Table 4.1 gives the return data series.
Excess returns for MSFT and the S &P 500 index are computed by
subtracting the Tbill rate from their monthly returns.
Table 4.1 Monthly returns for Microsoft (MSFT), S &P 500 index, and Treasury bills from January
2018 to December 2020

Year/Month MSFT Tbill S &P 500 Excess MSFT Excess S &P 500
18/1 −1.31 0.11 -3.89 -1.41 -4.00
18/2 −2.67 0.12 −2.69 −2.78 −2.80
18/3 2.47 0.14 0.27 2.33 0.14
18/4 5.69 0.14 2.16 5.55 2.02
18/5 −0.23 0.14 0.48 −0.37 0.34
18/6 7.58 0.15 3.60 7.42 3.45
18/7 5.89 0.16 3.03 5.73 2.87
18/8 1.82 0.16 0.43 1.65 0.27
18/9 −6.61 0.17 −6.94 −6.78 −7.11
18/10 3.82 0.18 1.79 3.64 1.60
18/11 −8.40 0.19 −9.18 −8.59 −9.36
18/12 2.82 0.20 7.87 2.62 7.67
19/1 7.28 0.20 2.97 7.08 2.77
19/2 5.28 0.20 1.79 5.07 1.59
19/3 10.73 0.20 3.93 10.53 3.73
19/4 −5.30 0.20 −6.58 −5.50 −6.78
19/5 8.31 0.20 6.89 8.11 6.69
19/6 1.72 0.19 1.31 1.54 1.13
19/7 1.17 0.18 −1.81 0.99 −1.99
Year/Month MSFT Tbill S &P 500 Excess MSFT Excess S &P 500
19/8 0.85 0.17 1.72 0.68 1.55
19/9 3.12 0.17 2.04 2.96 1.88
19/10 5.59 0.14 3.40 5.44 3.26
19/11 4.17 0.13 2.86 4.04 2.73
19/12 7.95 0.13 −0.16 7.82 −0.29
20/1 −4.83 0.13 −8.41 −4.96 −8.54
20/2 −2.65 0.13 −12.51 −2.79 −12.64
20/3 13.63 0.03 12.68 13.60 12.65
20/4 2.25 0.01 4.53 2.25 4.52
20/5 11.06 0.01 1.84 11.05 1.83
20/6 0.74 0.01 5.51 0.73 5.50
20/7 10.01 0.01 7.01 10.00 7.00
20/8 −6.74 0.01 −3.92 −6.75 −3.93
20/9 −3.74 0.01 −2.77 −3.74 −2.77
20/10 5.73 0.01 10.75 5.72 10.75
20/11 3.90 0.01 3.71 3.89 3.70
20/12 −2.22 0.01 1.43 −2.22 1.43

This table reports the monthly returns (in percent) for Microsoft (MSFT),
the S &P 500 index, and Treasury bills from January 2018 to December
2020. Stock prices are downloaded from Yahoo Finance. Adjusted stock
prices are used that take into account dividends and splits. Treasury bill
rates are downloaded from the FRED database provided online by the
Federal Reserve Bank of St. Louis (see https://fred.stlouisfed.org/series/
DGS1MO)

To implement the market model, we estimate the simple regression Eq.

(4.3). Using our variables, we can re-write this equation as:
$$\begin{aligned} R_{ MSFT t}-R_{ft} = \hat{\alpha }_{ MSFT }
+ \hat{\beta }_{ MSFT }(R_{ S \&amp; P500 t}-R_{ft}) + e_{ MSFT (4.6)
t}. \end{aligned}$$
In Excel, open the Data tab, then select Data Analysis. From the dropdown
menu select Regression. Enter the Input Y Range (excess MSFT return) and
Input X Range (excess S &P 500 return). You can do this by blocking off
the data column values you want to enter into the regression analysis. As
you block off the data column, you will see the Y or X Range values
change. Finally, select the Output Options that you want, and click the OK
button. Let’s look at some of the output of the market model for MSFT.

Fig. 4.1 The estimated market model for Microsoft (MSFT) stock returns with fitted regression line

Figure 4.1 plots the excess returns of MSFT (or Y variable) and the S
&P 500 index (or X-variable) as well as the fitted regression line. The fitted
line is comprised of the predicted values of the market model. The predicted
MSFT return for each monthly S &P 500 return is computed as:
$$\begin{aligned} R_{ MSFT t}-R_{ft}=\hat{\alpha }_{ MSFT
}+\hat{\beta }_{ MSFT }(R_{ S \&amp; P500 t}-R_{ft}), (4.7)
\end{aligned}$$
where $$\hat{\alpha }_{ MSFT }$$ is the estimated intercept
parameter, and $$\hat{\beta }_{ MSFT }$$ is the estimated slope
measuring the beta risk of MSFT. We provide a screenshot of the Excel
output below the diagram. The fitted line plots the predicted values of
excess MSFT returns (see faded points on the straight line). The monthly
returns for MSFT are represented by the scatter plot of points around the
fitted line. The difference between each scatter plot point and the fitted line
equals the residual $$e_{ MSFT ,t}$$ in Eq. (4.6). The larger the
residuals, the lower the goodness-of-fit of the market model. In this regard,
the Excel output reports the commonly used goodness-of-fit measure
known as the adjusted R-squared . In the present case, MSFT had an
estimated R-squared value of 61.4%, which means that the fit is fairly good.
The $$\alpha$$ estimate is greater than zero at
$$\hat{\alpha }_{ MSFT } = 1.67$$, and the $$\beta$$ estimate
is less than one at $$\hat{\beta }_{ MSFT } =0.79$$. The t-statistics
for these estimated parameters are 2.88 and 7.35, respectively, which have
very low p-values below 0.01 suggesting at least a 1% level of significance.
These results indicate that $$\hat{\alpha }_{ MSFT }$$ and
$$\hat{\beta }_{ MSFT }$$ are highly significant in terms of their
difference from zero. The positive and highly significant
 $$\hat{\alpha }$$ implies that other factors are needed to fully price
MSFT. Also, the positive and even more highly significant
$$\hat{\beta }_{ MSFT }$$ can be interpreted to mean that the excess
market factor proxied by S &P 500 index returns minus Tbill rates helps to
explain movements in MSFT excess returns in our sample period. Because
this estimate is less than one, we can infer that MSFT has less systematic
risk than the market as a whole (or at least the largest 500 stocks).
The market model results for MSFT are consistent with the CAPM for
the most part but not entirely. There appears to be a linear relation between
its excess stock returns and excess market returns in line with the CAPM.
However, according to the CAPM, the intercept or $$\alpha$$
parameter should be zero. Clearly, the large $$\alpha$$ term at 1.67%
per month suggests that a large part of MSFT’s excess returns are not
explained by excess market returns. For comparison purposes, the average
excess MSFT return and average excess market return in the sample period
were 2.35% and 0.86% per month, respectively. Hence, 1.67% is quite large
in terms of mispricing error captured by the $$\alpha$$ term. It
appears that some additional factors are needed to more fully price MSFT
and thereby reduce the magnitude of $$\alpha$$ mispricing.

4.4 A Primer on Regression Analysis

This section covers some fundamentals of regression models. Regression
techniques are arguably the most commonly used empirical tools in
economics and finance. The underlying idea is to empirically capture the
relation between independent explanatory variables (x) and dependent
variables (y). A simple regression with one explanatory variable is:
$$\begin{aligned} y_i = \alpha + \beta x_i + e_i, \end{aligned}$$ (4.8)
where $$i = 1, \ldots , n$$ is the $$i^{\text {th}}$$ observation for a
sample of size n observations, $$\alpha$$ is the intercept term (or
constant term), $$\beta$$ the slope coefficient, and e is the unobserved
error term. Note that $$\alpha$$ and $$\beta$$ are just generic
notations for the regression coefficients and are not related to the CAPM
discussed earlier. The task is to estimate the unknown regression
parameters $$\alpha$$ and $$\beta$$ based on a sample of n observed
value pairs $$(x_i, y_i)$$, $$i = 1, \ldots , n$$.
The setup in Eq. (4.8) can be readily extended to multiple regression
with k explanatory variables, or $$x_1, \ldots , x_k$$,
$$\begin{aligned} y_i = \alpha + \beta _1 x_{i1} + \cdots + \beta _k
(4.9)
x_{ik} + e_i, \end{aligned}$$
where $$\alpha$$ is the intercept, $$\beta _j$$ are slope
coefficients for $$j = 1, \ldots , k$$ independent variables, and e is the
unobserved error term. Now the estimation of the coefficients is based on n
observations on the variables $$(x_ 1, \ldots , x_k, y)$$ with the
$$i^{\text {th}}$$ observation denoted as
$$(x_{i1}, x_{i2}, \ldots , x_{ik}, y_i)$$. Hence, the sample of n
observations forms an $$n\times k$$ matrix similar to a data table in
Excel with variable names occupying the first row of $$p+1$$
columns and the n rows containing the observed values of the k x-variables
and the dependent variable y at each time t.
The above simple regression in (4.8) and multiple regression (4.9) are
far more general than might be thought at first glance. For example, we can
freely make transformations of the variables. For example, the logarithmic
transformation is very popular in economics to transform absolute values
into relative values. Importantly, fairly general non-linear relationships
between the dependent variable and independent variable can be dealt with
in the framework of the above linear setup. For example, polynomial
extensions of individual variables are often utilized to capture possible non-
linearities. Even so, the resulting equation can be represented by the above
linear form. The second pass cross-sectional FM regression in Eq. (4.5) is a
good example with $$y_{p} = R_{pt}$$,
$$x_{p,1} = \beta _{pt-1}$$, $$x_{p, 2} = \beta _{pt-1}^2$$,
and $$x_{p, 3} = \bar{\sigma }(e)_{pt-1}$$. Here the second
explanatory variable, or $$x_{p, 2}$$, is the square of the first
variable—that is, $$x_{p, 2} = x_{p, 1}^2$$ in the regression
equation.

4.4.1 On the Interpretation of Regression Coefficients

The properties of error term, or e, play a salient role in the statistical
estimation of regression parameters. The expected value of the error terms
is zero, i.e., $$E(e) = 0$$. This result can always be technically achieved
in regression with an intercept term. A crucial assumption is that the error
term is not correlated with the x variables. In particular, it is assumed that
$$E(e|x) = 0$$, which says that the conditional expectation of the error
term given x is zero. This property implies both zero correlation with x
variables as well as $$E(e) = 0$$. Given this assumption,
$$E(y | x) = \alpha + \beta x$$ for the simple regression and
$$E(y | x_1, \ldots , x_k) = \alpha + \beta _1 x_1 + \cdots + \beta _k
x_k$$
in the multiple regression, so that we have straightforward interpretations of
the regression coefficients. In the simple regression $$\beta$$ indicates
the average change in y when x changes by one unit. For example, in the
market model in Eq. (4.3), one percentage point change in the excess
market return $$R_{mt} - R_{ft}$$ is expected to change the excess
stock return $$R_{it} - R_{ft}$$ by $$\beta _i$$ percentage points.
The interpretation of multiple regression coefficients $$\beta _j$$
is similar. They indicate the (average) marginal effects of the $$x_j$$
s on y. That is, $$\beta _j$$ indicates the expected change in y when
$$x_j$$ changes by one unit given that other x-variables remain
unchanged. This interpretation assumes that $$x_j$$ is not
functionally related to the other x variables (i.e., the independent
variables are uncorrelated with one another).
In mathematical terms, the slope coefficients capture marginal effects in
that they are partial derivatives of the dependent variable with respect to the
explanatory variable. In more complicated cases like the FM multiple
regression in (4.5), the marginal effect of $$\beta _{pt-1}$$ is not
constant $$\lambda _{m}$$ but
$$\lambda _{m} + 2\lambda _{nl} \beta _{pt-1}$$. That is, the
strength of the marginal effect of $$\beta _{pt-1}$$ depends on its
level. If the slope coefficient $$\lambda _{nl}$$ were nonzero, it
would indicate a non-linear relation, which contradicts the CAPM but
reveals additional economic information about return-risk relationships.
Positive $$\lambda _{nl}$$ would imply that investors demand an
incremental premium for stocks with higher market risk. A negative
 $$\lambda _{nl}$$ implies the opposite. Hence, the interpretation of
regression coefficients must be based on the functional form of the initial
model.
In general, the intercept term is of lesser interest in most fields of study.
It is used to improve the goodness-of-fit of the regression model and satisfy
the OLS condition that $$E(e) = 0$$. In simple language, it captures
the value of the dependent variable not explained by the independent
variables. However, in asset pricing the intercept term has more than a
statistical interpretation. It plays a central role in measuring mispricing by
asset pricing models like the CAPM, as discussed in the previous section.
Indeed, Jensen’s alpha is routinely reported alongside beta in many
financial publications. From this perspective, mispricing alpha is an
important metric for evaluating the validity of asset pricing models.

4.4.2 Justification for OLS

Statistical theory (related to the Gauss-Markov Theorem) shows that under
certain assumptions OLS is the best linear unbiased estimator (BLUE) of
the regression coefficients. This means that among linear estimators, the
OLS estimator is unbiased and has smaller variance than any other linear
unbiased estimator. Unbiasedness is a desirable property that guarantees the
estimator does not systematically understate or overstate the population
parameter. In this regard, the variance measures the precision of an
estimator, wherein smaller is better. The underlying classical assumptions
are itemized below.
i.
Zero mean error terms: $$E(e_i) = 0$$ for all i.
ii.
Homoskedasticity (constant variance of error terms):
$$Var \,(e_i) = \sigma _e^2$$ for all i.
iii.
Errors are uncorrelated: $$Cov \,(e_i, e_j) = 0$$ for all $$i \ne j$$
.
iv.
No perfect multicollinearity: x-variables are not linearly dependent.
v.
$$Cov \,(x_i, e_i) = 0$$.
 Under these assumptions with $$n &gt; k + 1$$ sample
observations, the solution to the minimum values of error terms in the
regression equation or
$$\begin{aligned} (\hat{\alpha }, \hat{\beta }_1, \ldots , \hat{\beta
}_k) = \text {argmin}_{\alpha , \beta _1, \ldots , \beta _k} \sum _{i =
(4.10)
1}^n\left[ y_i - (\alpha + \beta _1 x_{i1} + \cdots + \beta _k
x_{ik})\right] ^2 \end{aligned}$$
gives the OLS estimators with the BLUE property.
In the case of the simple regression, the explicit solution is
$$\begin{aligned} \hat{\beta }= \frac{\sum _{i = 1}^n(x_i -
\bar{x})(y_i - \bar{y})}{\sum _{i = 1}^n(x_i - \bar{x})^2} (4.11)
\end{aligned}$$
and
$$\begin{aligned} \hat{\alpha }= \bar{y} - \hat{\beta }\bar{x},
(4.12)
\end{aligned}$$
where $$\bar{x}$$ and $$\bar{y}$$ are the sample means of x
and y, respectively.
For the multiple regression, it is more convenient to use matrix
notations and compile the y observations to a column vector
$$\textbf{y}= (y_1, \ldots , y_n)'$$ and the x-observations to an
$$\textbf{X}$$ matrix of the form
$$\begin{aligned} {\textbf{X}} = \left( \begin{array}{cccc} 1
&amp;{} x_{11} &amp;{} \cdots &amp;{} x_{1k} \\ 1 &amp;{}
x_{21} &amp;{} \cdots &amp;{} x_{2k} \\ \vdots &amp;{} \vdots (4.13)
&amp;{} &amp;{} \vdots \\ 1 &amp;{} x_{n1} &amp;{} \cdots
&amp;{} x_{nk} \end{array} \right) \end{aligned}$$
or $$n \times (k + 1)$$ matrix. Using matrix notations, the multiple
regression can be written compactly as
$$\begin{aligned} {\textbf{y}} = {\textbf{X}}{\boldsymbol{\beta
(4.14)
}} + {\textbf{e}}, \end{aligned}$$
where $$\boldsymbol{\beta }= (\alpha , \beta _1, \ldots , \beta _k)'$$
is the $$k + 1$$-vector of regression coefficients, and
 $$\textbf{e}= (e_1, \ldots , e_n)'$$ is the error term vector. Thus, the
first column of constants equal to one in $$\textbf{X}$$ has intercept
$$\alpha$$ as the slope coefficient, which is often called the
coefficient of the constant term. With these matrix notations, the OLS
estimator of $$\boldsymbol{\beta }$$ becomes
$$\begin{aligned} \hat{\boldsymbol{\beta }} =
(4.15)
(\textbf{X}'\textbf{X})^{-1}\textbf{X}' \textbf{y}. \end{aligned}$$
The above classical assumptions guarantee desirable properties for OLS
estimators. The properties even strengthen if we can assume that the errors
are normally distributed, i.e., $$e_i \sim N(0, \sigma _e^2)$$. Then
OLS also satisfies the properties of minimum variance unbiased estimators
(MVUEs)—that is, OLS produces the best estimators among all unbiased
estimators (not only among linear unbiased estimators).
Before continuing further, it is worthwhile to discuss the classical
assumptions. Assumption (i) on the zero expected value of error terms is
mechanical in nature as it holds whenever the intercept term
$$\alpha$$ is freely estimated. Thus, intercept terms are routinely
introduced in regression models. Reasonable justification is needed to
exclude intercept terms from a model. By excluding an intercept, the slope
coefficients may become biased.8
Assumption (ii) concerning homoskedasticity is technically desirable, as
violation of this assumption causes OLS to lose the “best” in the
BLUE qualities, i.e., best precision. The major problem is that the standard
errors of the coefficient estimators become typically biased downward due
to heteroskedasticity, thereby biasing inferences (e.g., over rejection of the
null hypothesis in testing). However, under fairly general conditions, OLS
is still consistent, meaning that as the sample size increases, the OLS
estimator converges toward the corresponding population slope coefficient.
By correcting the biases in standard errors with robust estimators,
White (1980) heteroskedastic robust variance estimators (denoted HC) are
nowadays routinely used as a solution to this potential problem.
Similarly, if the zero correlation of error terms assumption in (iii) is
violated, robust alternatives are routinely used to help correct the problem.
In time-series data, the solution is to use Newey and West (1987)
heteroskedasticity and autocorrelation robust standard errors (denoted
HAC). In cross-sectional observations, correlation of error terms can occur
if observations form clusters (like industries) within which observations
tend to be more alike than between clusters. In this situation, Cameron et al.
(2011) have proposed clustering robust standard errors as a solution.
The fourth assumption (iv) is a technical problem that can lead to
unstable regression parameter estimates. For example, a regression
coefficient could flip from positive to negative in sign due to correlation of
its independent variable with another independent variable in a multiple
regression model. Unlike macroeconomics, multicollinearity is not a
common problem in financial economics. Independent variables are
normally not highly correlated with one another.
The last assumption (v) of zero correlation between the error terms and
the explanatory variables is a serious and difficult problem that is more or
less violated in virtually all empirical estimations. The simple reason is that
the error term contains all the missing factors that affect the dependent
variable. In economics, independent variables tend to be correlated.
Therefore, missing factors absorbed into the error term can be correlated
with the explanatory variables. This complication is known as an
endogeneity problem that biases the slope coefficient estimates. Economic
variables that are typically positively correlated imply positive bias—that
is, over-estimation of the coefficients. Various instrumental
variable techniques have been developed to reduce the bias. However,
finding appropriate instruments is a difficult practical issue. In some cases,
panel data analysis can alleviate the problem. Unfortunately, there is no
easy solution to this problem. Relevant to our focus on asset pricing models
in finance, this problem is not generally serious.

4.5 OLS Estimation of the Market Model

We next apply the above discussion of OLS assumptions to show that the
market model in Eq. (4.3) is indeed the empirical counterpart of the CAPM
in Eq. (4.1). Under the CAPM, the classical Gauss–Markov assumptions are
satisfied to justify OLS estimation of the market model using time series of
stock returns. We begin by re-writing CAPM Eq. (4.1) with time index
(assuming for simplicity that the riskless rate is a constant $$R_f$$) as:
$$\begin{aligned} E(R_{it}) - R_{f} = \beta _i \left[ E(R_{Mt}) -
R_{f}\right] , \end{aligned}$$
Next, adding and subtracting $$R_{it}$$ on the left-hand-side, and
similarly adding and subtracting $$\beta _{i} R_{Mt}$$ on the right-
hand-side, after rearranging terms in the equation, we get:
$$\begin{aligned} R_{it} - R_f + \left[ E(R_{it}) - R_{it}\right] = \beta
_i(R_{Mt} - R_f) - \beta _i\left[ R_{Mt} - E(R_{Mt})\right] .
\end{aligned}$$
Moving $$\left[ E(R_{it}) - R_{it}\right]$$ to the right-hand-side and
simplifying terms, we can write the market model:
$$\begin{aligned} R_{it} - R_f = \beta _i(R_{Mt} - R_f) + u_{it},
(4.16)
\end{aligned}$$
where
$$u_{it} = \left[ R_{it} - E(R_{it})\right] - \beta _i \left[ R_{Mt} -
E(R_{Mt})\right]$$
. Now if the CAPM holds, we can use this equation to estimate the
unknown beta coefficient $$\beta _i$$ with OLS because all the classical
assumptions hold. We leave the details for the student to check that: (i)
$$E(u_{it}) = 0$$, (ii)
$$Var \,(u_{it}) = \sigma _i^2(1 - \rho _{iM}^2) \equiv \sigma
_{u_i}^2$$
is constant, wherein $$\rho _{iM} = \text {cor}\,(R_{it}, R_{Mt})$$,9
and (iii) $$Cov \,(u_{it}, u_{is}) = 0$$ for all $$t \ne s$$ due to being
serially uncorrelated in an efficient market, which implies that
$$Cov \,(R_{it}, R_{is}) = 0$$, $$Cov \,(R_{Mt}, R_{Ms}) = 0$$,
and $$Cov \,(R_{it}, R_{Ms}) = 0$$ for all $$t\ne s$$. The condition
(iv) can be satisfied directly by the researcher by choosing uncorrelated
variables (i.e., no multicollinearity). Finally, condition (v) is satisfied due to
properties of the covariance whereby
$$Cov \,(u_{it}, u_{is}) = Cov (R_{it}, R_{is}) - \beta _i Cov \,(R_{it},
R_{Ms}) + \beta _i^2 Cov \,(R_{it}, R_{Ms})$$
in which all covariances are zero due to market efficiency whenever
$$t\ne s$$. Consequently, we know that $$Cov \,(u_{it}, u_{is}) = 0$$
for all $$t \ne s$$. Due to satisfying these assumptions, the beta
parameter can be efficiently estimated by the market model.
 However, if the CAPM does not hold, and there are additional factors
affecting stock returns, the situation changes. The missing factors will be
absorbed into the error term $$u_{it}$$. If the factors are not
correlated with the market return and only affect the mean return, such that
$$E(u_{it}) = \alpha _i \ne 0$$, we can easily fix the problem by
redefining the error term as $$e_{it} = u_{it} - \alpha _i$$ and
balancing the subtraction by adding $$\alpha _i$$ into Eq. (4.16). The
end result is Eq. (4.3), or
$$\begin{aligned} R_{it} - R_f = \alpha _i + \beta _i(R_{Mt} - R_f) +
e_{it}. \end{aligned}$$
Again all of the OLS assumptions are satisfied, and OLS can be used to
produce efficient estimates (BLUE) for the parameters assuming the
unknown factors are not (materially) correlated with the market factor. Is
the beta estimate in the market model significant? Appendix A provides a
discussion related to statistical inference from sampling theory, which
enables tests of a beta estimate to see whether it is statistically significant or
not.

4.5.1 Goodness-of-Fit
The most popular goodness-of-fit measure in regression analysis is R-
squared (also known as the coefficient of determination) and its adjusted
variant often denoted as $$\bar{R}$$-squared. These measures tell us
how well the regression equation fits the data observations. The basic idea
is that, if the regression equation yields predicted values that are close to the
actual values of observations, it has strong goodness-of-fit, and vice versa if
predicted and actual values are very different from one another. To compute
goodness-of-fit, we need to decompose the total variation measure
$$\text {SST} = \sum _{i = 1}^n(y_i - \bar{y})^2$$ of the dependent
variable. Letting
$$\hat{y}_i = \hat{\alpha }+ \hat{\beta }_1x_{i1} + \cdots + \hat{\beta
}_kx_{ik}$$
denote the fitted value, the decomposition becomes
$$\begin{aligned} \sum _{i = 1}^n(y_i - \bar{y})^2 = \sum _{i =
1}^n(\hat{y}_i - \bar{y})^2 + \sum _{i = 1}^n(y_i - \hat{y}_i)^2, (4.17)
\end{aligned}$$
or
$$\begin{aligned} \text {SST} = \text {SSE} + \text {SSR}
(4.18)
\end{aligned}$$
where $$\text {SST} = \sum _{i = 1}^n(y_i - \bar{y})^2$$ is the total
sum of squares,
$$\text {SSE} = \sum _{i = 1}^n(\hat{y}_i - \bar{y})^2$$ is the sum of
squares explained by the regression which measures the share of variability
around the mean $$\bar{y}$$ that is due to the independent variables in
the regression model, and
$$\text {SSR} = \sum _{i = 1}^n(y_i - \hat{y}_i)^2$$ is the sum of
squares due to residuals which is the variation not captured by the estimated
regression model. The R-squared is the ratio of the total variation explained
by the regression model,
$$R^{2} = \frac{{{\text{SSE}}}}{{{\text{SST}}}} = 1 -
(4.19)
\frac{{{\text{SSR}}}}{{{\text{SST}}}}.$$
The adjusted R-squared is based on the rightmost representation by
adjusting the sums of squares using the respective degrees of freedom with
k independent variables such that
$$\begin{aligned} \bar{R}^2 = 1 - \frac{s_e^2}{s_y^2},
(4.20)
\end{aligned}$$
where
$$\begin{aligned} s_e^2 = \frac{1}{n - k - 1}\sum _{i = 1}^n(y_i -
(4.21)
\hat{y}_i)^2 \end{aligned}$$
is the unbiased OLS variance estimator of the error variance
$$\sigma _e^2$$, and $$s_y^2$$ is the usual unbiased estimator of the
variance of y.
The major difference between $$R^2$$ and $$\bar{R}^2$$
is that the latter measure is penalized by the degrees of freedom. While
adding a new variable can reduce $$\bar{R}^2$$ if the variable does
not improve the explanatory power enough, the non-adjusted
$$R^2$$ tends to increase or at least never decrease. By definition
$$0 \le R^2 \le 1$$. Note that the adjusted R-squared can be negative
if none of the explanatory variables helps to predict y. For these reasons, the
adjusted R-squared is more popular to report, as a positive value indicates
that the fitted model has some explanatory power.

4.6 Summary
The empirical form of the CAPM named the market model was proposed by
Markowitz (1959) and developed by Sharpe (1963) and Fama (1968).
Excess asset returns (dependent y variable) are regressed on excess market
portfolio proxy returns (independent x variable) to yield an $$\alpha$$
intercept parameter and beta $$\beta$$ slope parameter. Jensen (1968)
recognized that the $$\alpha$$ terms, so-called Jensen’s alpha, can be
used to jointly test whether the proxy for the market portfolio is efficient or
mispricing exists due to missing factors. According to the CAPM, the
$$\alpha$$ estimate should equal zero in the market model. In the real
world, stocks and other assets do not lie exactly on the Security Market
Line (SML), which results in a random error term in the market model.
Early tests of the CAPM employed the market model. Black, Jensen,
and Scholes (BJS) (1972) found that the SML was flatter than predicted by
the CAPM. They formed 10 portfolios of U.S. stocks sorted by beta.
Contrary to the CAPM, 3-out-of-10 portfolios had significant
$$\alpha$$ intercept terms. Also, stock portfolios with betas greater
(less) than one tended to have negative (positive) intercepts. These results
suggested that low (high) beta stocks had higher (lower) returns than
predicted by the CAPM. Additionally, average $$\alpha$$s for their
stock portfolios were significantly different from zero. This finding means
that either the market portfolio proxy is inefficient or missing factors exist,
which rejects the CAPM.
Another study by Fama and MacBeth (FM) (1973) conducted further
market model tests using U.S. stock portfolio returns. They proposed a two-
step estimation procedure: (1) a time-series regression model is used to
estimate model parameters; and (2) a cross-sectional regression is used to
estimate the market price of risk (or $$\lambda$$ coefficient)
associated with the step (1) estimated model parameters (called loadings).
Some key findings in their study were: (1) time-series regressions showed
that non-linear beta and non-beta risks were not significant, only beta
risk was significant; and (2) cross-sectional regression showed that
 $$\lambda _m$$ associated with beta risk was significant meaning
that beta risk is priced in the cross-section of stock returns. Because the
market price of beta risk tended to be less than the actual excess return of
the market portfolio proxy, they inferred that the borrowing rate is greater
than the riskless government bond rate.
Understanding the fundamentals of regression analyses is crucial to
asset pricing. Simple regression with one independent variable and multiple
regression with multiple independent variables are important tools in
finance. Both regression models can be adapted to different transformations
of the variables, including logarithms of variables, non-linear variables, etc.
Regression $$\beta$$ coefficients measure the average change in y
when x changes by one unit, given other independent variables are held
constant in the model. The intercept $$\alpha$$ coefficient is used in
regression models to improve the goodness-of-fit and ensure that the
expected value of the error terms (denoted e) equals zero.
Under the Gauss–Markov Theorem, ordinary least squares (OLS)
provide the best linear unbiased estimator (BLUE) of the regression
coefficient. This theorem requires a number of statistical properties—(i)
zero mean error terms: $$E(e_i) = 0$$ for all i; (ii)
homoskedasticity (constant variance of error terms):
$$Var \,(e_i) = \sigma _e^2$$ for all i; (iii) errors are uncorrelated:
$$Cov \,(e_i, e_j) = 0$$ for all $$i \ne j$$; (iv) no perfect
multicollinearity: x-variables are not linearly dependent; and (v)
$$Cov \,(x_i, e_i) = 0$$. Assumption (i) is achieved by using an
intercept term. If assumption (ii) is violated due to heteroskedaisticity (or
nonconstant error variance), a White (1980) estimator (HC) can be used to
adjust for this potential bias. If assumption (iii) is violated due to correlation
of errors in a time-series regression, a Newey and West (1987) estimator
(HAC) can be used. For cross-sectional regression, an adjustment by
Cameron, Gelbach, and Miller (2011) can be implemented. The no
multicollinearity assumption (iv) concerns the correlation between multiple
independent variables, which normally is not a major problem in financial
economics. Lastly, assumption (v) requires zero correlation between the
error terms and independent variables. If nonzero, an endogeneity
problem exists. This problem can be treated by introducing instrumental
variables into the regression model but is not easily accomplished.
 Interestingly, under the OLS assumptions, it can be proven that the
market model represents an empirical counterpart of the theoretical CAPM.
That is, due to satisfying these assumptions, the market model can
efficiently estimate the beta parameter. To test whether beta is significantly
different from zero, a t-ratio can be computed using the beta estimate and
the standard error of this estimate. In the case that the CAPM does not hold,
an intercept term is needed in the market model. This form of the market
model is typically used in empirical studies. Finally, the goodness-of-fit of
regression models is measured by R-squared measures. An adjusted
$$\bar{R}$$-squared is normally used to adjust for the number of
degrees of freedom that is affected by the number of regression parameters
in the model. In words, this goodness-of-fit statistic measures the amount of
variation around the mean value of the dependent variable explained by the
independent variables in the regression model.

Questions
1.
Write the equation for the market model form of the CAPM. What is
the intercept term in this model? Can the intercept be used to test the
CAPM?
2.
Why put a random error term in the market model? How would you
estimate the market model with regression analysis?
3.
Black, Jensen, and Scholes (BJS) (1972) tested the CAPM using the
market model. Why did they use portfolios rather than individual
stocks in their tests? How were the portfolios formed for their tests?
Write the t-test that they performed to test whether the
$$\alpha _p$$ for a portfolio was not equal to zero. What were the
three main findings of their study? What did they conclude about the
CAPM?
4.
Fama and MacBeth (FM) (1973) tested the CAPM using cross-
sectional regression analysis. What is the cross-sectional regression
model that they tested?
5. What did Fama and MacBeth (1973) find in their cross-sectional tests
of the CAPM?
6.
Write a simple regression model with one explanatory variable. Also,
write a multiple regression model with k explanatory variables
$$x_1, \ldots , x_k$$. How would you test for a non-linear
relationship between an independent variable and the dependent
variable?
7.
What are the statistical properties of the error term, or e, in a
regression model? How should the regression coefficient $$\beta$$
be interpreted? How should the intercept $$\alpha$$ be interpreted
from both a statistical perspective and an asset pricing perspective?
8.
List the underlying classical assumptions in OLS regression analysis.
Why are these assumptions important to OLS estimation of regression
coefficients?
9.
What does homoskedasticity of error terms mean in regression
analysis? Are error terms correlated with one another? What is
multicollinearity and can it cause problems in a multiple regression
model? Lastly, what is the problem if error terms in the regression
model are correlated with an independent variable?
10.
If the CAPM does not hold, how does this affect the market model?
11.
You want to test whether an estimated $$\beta _j$$ coefficient in an
OLS regression model is significant. What is the null hypothesis?
What is the statistical test of this hypothesis? What values of the t-
statistic are needed to reach 5% and 1% significance levels? (Hint:
this question is based on Appendix A.)
12.
What is goodness-of-fit mean in regression analysis? How would you
estimate it? Show a formula and explain its terms. Can it be adjusted
for the degrees of freedom in the estimation of the regression
equation? Which measure of goodness-of-fit should be used?
13.
You want to test whether the market risk of a stock is the same as the
market portfolio in CAPM. What is the null hypothesis? What is the
alternative hypothesis?
 x y
4.6 10.6
11.1 22.6
19.5 29.7
3.2 2.6
9.5 23.4
10.8 9.0
14.2 26.8
8.6 15.0
21.9 38.9
9.2 17.5

Compute the OLS estimates for the intercept and the slope coefficient for
the simple regression $$y = \alpha + \beta x + e$$ using Eqns. (4.11) and
(4.12).

Problems
1.
Update the data in Table 4.1 for Microsoft. Download monthly stock
returns for MSFT and the S &P 500 index from Yahoo Finance.
Download monthly Treasury bill rates from the Federal Reserve Bank
of St. Louis (https://fred.stlouisfed.org/series/DGS1MO). Gather data
for 36 months. What did you get for the intercept? Beta? Are these
model parameters significant as determined by t-statistics? Also, draw
a plot of the fitted regression line with excess MSFT returns on the Y-
axis and excess S &P 500 index returns on the X-axis. Do the results
support the CAPM?
2.
In the text, it is proved that OLS estimation of the market model is the
empirical counterpart of the CAPM under classical Gauss–Markov
assumptions. Show that: (i) $$E(u_{it}) = 0$$, (ii)
$$Var \,(u_{it}) = \sigma _i^2(1 - \rho _{iM}^2) \equiv \sigma
_{u_i}^2$$
is constant, and (iii) $$Cov \,(u_{it}, u_{is}) = 0$$ for all .
3. Consider the following observations for variables x and y:
Appendix A: Statistical Inference
Statistical inference gives us guidelines concerning what we can say about
the population slope coefficients on the basis of the sample estimates.
Confidence intervals and hypothesis testing are the major tools in this
process. Statistical properties of the error term $$e_{it}$$ are critical.
While the Gauss–Markov BLUE result needs only assumptions about the
expected values, variances, and covariances, statistical inference needs
more precise information about the whole distribution of the error terms.
The textbook assumption is that the error term is normally distributed,
or $$e_{i} \sim N(0, \sigma _{e}^2)$$. While this assumption
strengthens OLS estimator properties, as discussed in this chapter, it also
determines the distribution (called the sampling distribution) of the
estimator. An important property of the normal distribution is that any
linear combinations of normal random variables are again normally
distributed. Therefore, as an example considering $$\hat{\beta }$$ in
Eq. (4.10), using fairly straightforward calculations, we can write
$$\begin{aligned} \hat{\beta }= \beta + \sum _{i = 1}^n w_i e_i,
\end{aligned}$$
in which
$$w_{i} = (x_i - \bar{x}) / \sum _{i = 1}^n(x_i - \bar{x})^2$$, such
that $$\hat{\beta }_i$$ indeed is a linear combination of
$$e_i$$. Thus, given $$x_i$$, $$\hat{\beta }$$ is
normally distributed with expected value $$\beta$$ and variance
$$\begin{aligned} \sigma _{\hat{\beta }}^2 = \frac{\sigma _e^2}{\sum
_{i = 1}^n(x_i - \bar{x})^2}. \end{aligned}$$
In the same fashion, $$\hat{\alpha }$$ is normal with expected value
$$\alpha$$ and variance
$$\begin{aligned} \sigma _{\hat{\alpha }}^2 = \sigma _e^2\left( \frac{1}
{n} + \frac{\bar{x}^2}{\sum _{i = 1}^n(x_i - \bar{x})^2}\right) .
\end{aligned}$$
In the general case of Eq. (4.15), the variances of respective OLS
coefficients are obtained from the diagonal of the matrix
$$\begin{aligned} Var \,(\hat{\boldsymbol{\beta }}) = \sigma
(A4.1)
_e^2(\mathbf{X'X})^{-1}. \end{aligned}$$
If the error variance $$\sigma _e^2$$ were known, we could base our
inferences directly on the normal distribution. In practice, however,
variance $$\sigma _e^2$$ must be replaced by its regression estimate
$$s_e^2$$ in Eq. (4.21) with the implication that inferences are based
on the Student t-distribution. Thus, in a regression with k independent
variables, the t-ratio for testing the null hypothesis
$$\begin{aligned} H_0 : \beta _j = \beta _j^* \end{aligned}$$ (A4.2)
becomes
$$\begin{aligned} t = \frac{\hat{\beta }_j - \beta _j^*}
(A4.3)
{s_{\hat{\beta }_j}}, \end{aligned}$$
where $$\beta _j^*$$ is the hypothesized value of the population
coefficient (i.e., a fixed value defined by the researcher),
$$s_{\hat{\beta }_j}$$ is the estimated standard error of
$$\hat{\beta }_j$$, equaling the square root of the jth diagonal
element of Eq. (A4.1) with $$s_e^2$$ substituted for
$$\sigma _e^2$$, $$j = 0, 1, 2, \ldots , k$$, and
$$\beta _0$$ the intercept term (equaling $$\alpha$$ in the
text). Under the null hypothesis, the t-ratio in Eq. (A4.3) is t-distributed
with $$n - k - 1$$ degrees of freedom. Regarding the null hypothesis
in Eq. (A4.2), the most common case is that we are interested in whether
variable $$x_j$$ has any impact on y. In this case,
$$\beta _j^* = 0$$ in (A4.2) so that the null hypothesis becomes
$$\begin{aligned} H_0: \beta _j = 0 \end{aligned}$$
and the t-statistic in (A4.3) reduces to
$$\begin{aligned} t = \frac{\hat{\beta }_j}{s_{\hat{\beta }_j}},
\end{aligned}$$
which is the t-ratio shown in standard computer outputs.
 Confidence intervals are of the form
$$\hat{\beta }\pm t_ps_{\hat{\beta }}$$ in which $$t_p$$ is an
appropriate percentile from the t-distribution to match the desired
confidence interval. Typical values are $$p = .025$$ for 95% and
$$p = .005$$ for 99% interval. The $$t_p$$ value itself depends
on the degrees of freedom. Typically in financial applications the number of
observations is large. Therefore, as the t-distribution approaches the normal
distribution when the degrees of freedom increases,
$$t_p \approx z_p$$ such that for the 95% interval
$$t_{.025} = 1.96 \approx 2$$ and for the 99% interval
$$t_{.005} = 2.58$$. These t-statistics are interpreted as 5% and 1%
significance levels, respectively. If t-statistics reach these levels or higher,
they indicate that the estimated beta coefficient is statistically significant.
Notably, even if the coefficient is statistically significant, it may not be
economically significant. The magnitude of the beta coefficient needs to be
large enough to suggest that a change in the independent variable will result
in a economically meaningful change in the dependent variable.
If the normality of the error term does not hold, the Central Limit
Theorem (CLT) provides us with so-called large sample results under fairly
mild assumptions. The CLT implies that the above normal
distribution results can be reliably used when the sample size is reasonably
large. There is no exact definition of how large the sample should be. In
general, with as small as 50 observations, the approximation is fairly close
for practical purposes.
As discussed earlier in the chapter, if the error terms are heteroskedastic
and/or autocorrelated, the major implication is bias in the standard
errors leading in most cases to overconfidence about the precision of the
estimates. For cross-sectional data, the White (1980) heteroskedastic robust
variance estimators (HC) are routinely used to correct the bias. In time-
series regression, the Newey and West (1987) heteroskedastic
autocorrelation robust measure (HAC) is used to correct standard errors.
And, if data are clustered, Cameron, Gelbach, and Miller (2011) clustering
robust standard errors can be used to take into account the correlation of
error terms.
As an example, we apply HAC standard errors to our earlier example of
Microsoft CAPM estimates in Sect. 4.3. For this purpose, we need more
sophisticated regression software than provided by Excel. R (https://r-
project.org) is a powerful open source software for general statistical
analyses as well as a wide variety of regression estimation methods.
Below is R-code for the estimation in which we assume that the data is
in an Excel file named msft.xlsx with the above headings on the first
line.

We observe that the HAC standard is just slightly larger than those of
OLS, which indicates that there is no material heteroskedasticity and/or
autocorrelation in the OLS residuals. Therefore, the OLS results are equally
reliable as those produced by the HAC approach, and the conclusions of
Sect. 4.3 pass this robustness check.

References
Black, F. 1972. Capital market equilibrium with restricted borrowing. Journal of Business 45: 444–
454.
[Crossref]

Black, F., M.C. Jensen, and M. Scholes. 1972. The capital asset pricing model: Some empirical tests.
In Studies in the Theory of Capital Markets, ed. M.C. Jensen. New York, NY: Praeger.

Blume, M. 1970. Portfolio theory: A step towards its practical application. Journal of Business 43:
152–174.
[Crossref]

Cameron, A.C., J.B. Gelbach, and D.L. Miller. 2011. Robust inference with multiway clustering.
Journal of Business and Economic Statistics 29: 238–249.
[Crossref]

Cochrane, J.H. 2005. Asset Pricing. Revised. Princeton, NJ: Princeton University Press.

Douglas, G.W. 1969. Risk in the equity markets: An empirical appraisal of market efficiency. Yale
Economic Essays 9: 3–45.

Fama, E.F. 1968. Risk, return, and equilibrium: Some clarifying comments. Journal of Finance 23:
29–40.
[Crossref]

Fama, E.F., and J.D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy 81: 607–636.
[Crossref]
Friend, I., and M. Blume. 1970. Measurement of portfolio performance under uncertainty. American
Economic Review 60: 607–636.

Gibbons, M.R., and J. Shanken. 1983. A test of the Sharpe-Lintner CAPM. Working Paper,
University of California, Berkeley, CA.

Jensen, M.C. 1968. The performance of mutual funds in the period 1945–1964. Journal of Finance
23: 389–416.
[Crossref]

Lintner, J. 1965. The valuation of risk assets and the selection of risky investments in stock portfolios
and capital budgets, Review of Economics and Statistics 47: 13–37.
[Crossref]

Markowitz, H.M. 1959. Portfolio Selection: Efficient Diversification of Investments. New York, NY:
Wiley.

Merton, R.C. 1972. An analytical derivation of the efficient portfolio frontier. Journal of Financial
and Quantitative Finance 7: 1851–1872.
[Crossref]

Miller, M., and M. Scholes. 1972. Rates of return in relation to risk: A re-examination of some recent
findings. In Studies in the Theory of Capital Markets, ed. M.C. Jensen, 47–78. New York, NY:
Praeger.

Mossin, J. 1966. Equilibrium in a capital asset market. Econometrica 34: 768–783.

[Crossref]

Newey, W., and K. West. 1987. A simple, positive semi-definite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica 55: 703–708.
[Crossref]

Roll, R. 1977. A critique of the asset pricing theory’s tests, part I: On past and potential future
testability of the theory. Journal of Financial Economics 4: 129–176.
[Crossref]

Roll, R. 1980. Orthogonal portfolios. Journal of Financial and Quantitative Analysis 15: 1005–1012.
[Crossref]

Shanken, J. 1992. On the estimation of beta pricing models. Review of Financial Studies 5: 1–34.
[Crossref]

Sharpe, W.F. 1963. A simplified model for portfolio analysis. Management Science 9: 277–293.
[Crossref]

Sharpe, W.F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance 19: 425–442.

Stambaugh, R. F. 1982. On the exclusion of assets from tests of the the two-parameter model.
Journal of Financial Economics 10: 237–268.
[Crossref]
Treynor, J.L. 1961. Market value, time, and risk. Unpublished manuscript.

Treynor, J.L. 1962. Toward a theory of market value of risky assets. Unpublished manuscript.

White, H. 1980. A heteroskedasticity-consistent covariance matrix estimator and a direct test for
heteroskedasticity. Econometrica 22: 193–218.

Footnotes
1 See also Black et al. (1972).

2 Treasury rates were only available from 1948 to 1966. From 1925 to 1947 they used the dealer
commercial paper rate for the riskless rate proxy.

3 See Blume (1970) and Friend and Blume (1970).

4 In forthcoming Sect. 4.4, we discuss the error term and other OLS assumptions in more detail.

5 The exact details of their data are somewhat complicated and beyond the scope of the present
general discussion. See Fama and MacBeth (1973, pp. 616–617).

6 It is well known that this rolling FM procedure eliminates correlation between residuals, or , in
cross-sectional regressions that can bias average risk premium t-tests. Some researchers conduct the
cross-sectional regressions within the sample by regressing portfolio returns in each month in the
sample period on average estimates for the entire period. In this case, it is necessary to make a
correction to adjust the standard errors in the t-statistic. Excellent discussions of this correction can
be found in Shanken (1992) and Cochrane (2005, Chapter 12).

7 See https://fred.stlouisfed.org/series/DGS1MO.

8 In the estimation of the market model and other asset pricing models with multiple factors, the
authors have found that the beta coefficients are changed little if any. Nonetheless, inclusion of an
intercept is more appropriate to ensure that the regression coefficients are BLUE.
9 The term , where , , and
. In the CAPM, , and using
in above, we get after adding up and rearranging terms
..
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_5

5. The Zero-Beta CAPM

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_5

In its early years, as discussed in the last chapter, the CAPM came under
question. Empirical tests of the market model found that the relationship
between CAPM beta and U.S. stock returns was weaker than expected. The
estimated Security Market Line (SML) relating the excess returns of stock
portfolios on the Y-axis and their estimated betas on the X-axis was flatter
than predicted by the CAPM with a higher intercept (or $$\alpha$$
parameter). Black, Jensen, and Scholes (1972) found that low (high) beta
stocks had higher (lower) returns than theorized by the CAPM.
Consequently, they proposed that the borrowing rate in the real world is
greater than the riskless rate used in the theoretical CAPM.
With this empirical evidence in hand, Black (1972) created the zero-
beta CAPM by relaxing key assumptions in the CAPM: (1) investors can
invest in both long and short positions in risky assets; and (2) investors can
borrow in financial markets at interest rates greater than the riskless rate.
Applying these assumptions, Black derived an alternative form of the
CAPM. We next review of the zero-beta CAPM in addition to empirical
tests of this model.

5.1 Zero-Beta CAPM

5.1.1 Assumptions
Black (1972) motivated his paper by overviewing the CAPM assumptions:
(1) investors have homogeneous expectations or opinions about assets’
distributions of returns; (2) asset returns follow a normal distribution; (3)
risk-averse investors seek to maximize the end-of-period expected utility of
wealth; and (4) investors can only hold long positions with unlimited
borrowing and lending at the riskless asset. For interested readers, Fama
(1968) provides an in-depth discussion of these assumptions. Contrary to
the CAPM Black modified assumption (4) by allowing short sales of assets.
Also, he believed that borrowing at the riskless rate was unrealistic. In
support, he cited earlier evidence in Black et al. (1972) that showed the
following model provides a better fit to U.S. stock returns than the CAPM
market model:its zero-beta portfolio
$$\begin{aligned} R_{it}=a_i+\beta _i R_{mt}+(1-\beta _i)
(5.1)
R_{zt}+ e_{it}, \end{aligned}$$
where the return on market proxy portfolio m (denoted $$R_{mt}$$) and
the return on the zero-beta portfolioz (denoted $$R_{zt}$$) have zero
correlation. Notice that the riskless rate $$R_{ft}$$ in the CAPM market
model has been replaced with the risky rate $$R_{zt} &gt; R_{ft}$$.
Using this higher borrowing rate for investors helps to explain the higher
intercept (or $$\alpha$$ parameter) in empirical tests of the CAPM as
well as the lower slope of the SML. Upon relaxing short sale and borrowing
rate restrictions, Black derived an alternative form of the CAPM dubbed the
zero-beta CAPM.

5.1.2 Theoretical Model

Using the single-period framework of the CAPM, which enables us to drop
the time t subscript, Black’s zero-beta CAPM is specified as follows:
$$\begin{aligned} E({R}_i)=E({R}_{Z}) +\beta _{i}[E({R}_{M})-
(5.2)
E({R}_{Z})], \end{aligned}$$
where $$\beta _{i}$$ is the beta risk of asset i with respect to theoretical
market portfolio M, $$E({R}_M)$$ is the expected return on the market
portfolio, and $$E({R}_{Z})$$ is the expected return on the zero-beta
portfolio Z. This model differs from the CAPM with borrowing and lending
at $$R_f$$ instead of $$R_Z$$. The excess return of the market
portfolio is now $$E({R}_M)-E({R}_{Z})$$ instead of
$$E({R}_M)-R_{f}$$. This excess return represents the return on the
zero-investment portfolio with a long position in M and a short position in
Z. Assuming that $$E({R}_{Z})&gt;R_f$$, the slope of this equilibrium
relationship between expected excess returns and beta risk is lower than the
SML of the CAPM. Given a lower SML slope, the intercept becomes
higher than the theoretical CAPM. For interested readers, Appendix A
contains a rigorous mathematical derivation of the zero-beta CAPM.
While Black assumed that investors cannot borrow at the riskless rate, it
is possible for them to invest in this asset. Allowing investors to hold a
portfolio consisting of market portfolio M, zero-beta portfolio Z, and
riskless assets f, he re-wrote the zero-beta CAPM as follows:
$$\begin{aligned} E({R}_i)-R_f=\beta _{i,M}[E({R}_M)-
(5.3)
R_f]+\beta _{i,Z}[E({R}_Z)-R_f], \end{aligned}$$
where $$\beta _{i,M}$$ and $$\beta _{i,Z}$$ represent the beta
risks of asset i with respect to market portfolio M and zero-beta portfolio Z,
respectively. If the investor only buys market portfolio M, then
$$\beta _{i,M}=1$$ and $$\beta _{i,Z}=0$$ in Eq. (5.3). On the
other hand, if the investor only buys zero-beta portfolio Z, then
$$\beta _{i,M}=0$$ and $$\beta _{i,Z}=1$$.
In Black’s world, investors can hold different combinations of risky
portfolios M and Z that are efficient in the sense of maximum expected
return per unit variance of returns (or minimum variance for a given
expected return). If they combine any risky portfolio composed of M and Z
with the riskless asset, less risky efficient portfolios can be created. So there
are now two ways for the investor to hold efficient portfolios (not simply
combining the riskless asset and M as in the CAPM).

Fig. 5.1 A graphical depiction of Black’s zero-beta CAPM

Figure 5.1 shows a diagram of Black’s model.1 The risky zero-beta

portfolio rate is greater than the riskless rate, or
$$E({R}_{Z})&gt;R_f$$. The investor can buy (go long) at the
riskless rate and an efficient portfolio. The optimal efficient portfolio in this
case is the portfolio T at the tangent point to the efficient frontier. Using
Roll’s (1980) geometric approach, the market portfolio M can be used to
obtain the zero-beta portfolio rate as follows: (1) connect a ray between M
and G (the global minimum variance portfolio, (2) extend this ray to
intersect the Y-axis, and (3) from this point move horizontally to locate the
zero-beta portfolio Z. The investor has two efficient frontiers. The first part
is the straight line from the riskless rate $$R_f$$ to efficient portfolio
T. The second part is any efficient portfolio from portfolios T to S on the
upper part of the efficient frontier. All portfolios between T and S can be
constructed by the investor by means of combining portfolios M and Z with
different weights.
Imagine that the investor holds some risky efficient portfolio denoted I
that is a combination of M and Z. In this case, the zero-beta CAPM would
become2:
$$\begin{aligned} E({R}_i)-R_f=\beta _{i,I}[E({R}_I)-R_f]+\beta
(5.4)
_{i,ZI}[E({R}_{ZI})-R_f], \end{aligned}$$
where $$\beta _{i,I}$$ and $$\beta _{i,ZI}$$ become the beta
risks of asset i with respect to efficient portfolio I and zero-beta portfolio ZI,
respectively. Thus, as proven by Roll (1980), the market portfolio M and its
zero-beta portfolio ZM can be replaced by any efficient portfolio I and its
zero-beta portfolio ZI. Relevant to the CAPM, any two portfolios I and ZI
can be used to construct market portfolio M. It should be noted that
investors can use both long and short positions to construct portfolios I and
ZI. In the CAPM, only long positions can be used to construct market
portfolio M, which is a value-weighted index of all assets.3 However, many
studies have found that short positions are needed to compute mean-
variance efficient portfolios.4 Taken together, these concepts make clear
that the zero-beta CAPM is a more general equilibrium pricing model than
the CAPM that depends solely on long-only market portfolio M.

5.2 Empirical Tests

There are two ways to test for the existence of the zero-beta CAPM. To see
this, let’s return to the Fama and MacBeth (FM) (1973) study reviewed in
Chapter 4. FM proposed a two-step test procedure for the market model
form of the CAPM, which can be used to test any asset pricing model and
its factors. In the first step, the risk parameters of the model are estimated
for test asset portfolios (e.g., beta risk or $$\beta _p$$). In the second
step, the estimated risk parameters (or $$\beta _p$$s) are regressed on
one-month-ahead (or out-of-sample) returns. Referring back to Eq. (4.3) in
Chapter 4, the market model can be estimated in excess return form using
portfolios as follows:
$$\begin{aligned} R_{pt}-R_{ft} = \alpha _0+\beta _p (R_{mt}-
(5.5)
R_{ft})+e_{it}, \end{aligned}$$
where $$R_{pt} - R_{ft}$$ is the excess rate of return on portfolio p in
time period t (e.g., daily or monthly), and $$R_{mt} - R_{ft}$$ is the
realized excess average rate of return on the stock market as a whole. In the
market model, if the CAPM is correct, the intercept term $$\alpha _0$$
should equal zero. However, if Black’s zero-beta CAPM holds, then using
Eq. (5.3), the intercept should be
$$\alpha _0 = E({R}_{Z}) - R_f &gt; 0$$. So one way to test the zero-
beta CAPM is to determine if there is a positive $$\alpha _0$$ term in the
above market model.
Again referring back to Chapter 4, and dropping non-linear terms in our
earlier Eq. (4.5), a cross-sectional regression can be estimated using market
factor loadings (or $$\beta _p$$s for different portfolios) as follows:
$$\begin{aligned} R_{pt+1}-R_{ft+1}=\lambda _0+\lambda _{m}
(5.6)
\beta _{pt}+u_{pt+1}, \end{aligned}$$
where $$R_{pt+1}-R_{ft+1}$$ is the portfolio excess return in the
next out-of-sample month $$t+1$$, $$\lambda _0$$ is the
intercept term not explained by the other variables in the equation,
$$\beta _{pt}$$ is the portfolio beta estimated using monthly returns
in previous years (e.g., four or five years), and $$u_{pt+1}$$ are
independent and normally distributed error terms.
Cross-sectional regression (5.6) asks the following question: Do the
estimated risk parameters for the portfolios line up with their future excess
returns? The coefficient $$\lambda _{m}$$ is an estimate of the
market risk premium, or $$E({R}_{Mt})-R_{ft}$$. According to
Ferson (2019, p. 223), $$\lambda _{m}$$ can be interpreted as a
zero-investment portfolio return that is long high-beta assets and short low-
beta assets. Furthermore, the intercept $$\lambda _0$$ provides an
estimate of the expected risk premium on the zero-beta portfolio, or
$$E({R}_{Zt}) - R_{ft}$$, which will also be a zero-investment
long/short portfolio of assets.5 If the CAPM holds, then
$$E({R}_{Zt}) = R_{ft}$$ and $$\lambda _0$$ should equal
zero. Hence, a second way to test the zero-beta CAPM is to check whether
the $$\lambda _0$$ term is positive.
It is important to recognize that the intercept term $$\alpha _0$$
in the time-series regression (5.5) and intercept term $$\lambda _0$$
in cross-sectional regression (5.6) are very different from one another. Both
of these intercepts provide estimates of the expected risk premium on the
zero-beta portfolio, but their estimation procedures materially affect their
values. In Eq. (5.5) the intercept $$\alpha _0$$ is estimated in a
specific estimation period (e.g., five years of monthly excess returns). By
contrast, the intercept $$\lambda _0$$ in Eq. (5.6) is based on one-
month-ahead excess returns that are out-of-sample in the sense that they are
not part of the estimation period for Eq. (5.5).
Another difference in these intercept terms is that $$\alpha _0$$
in time-series regression (5.5) is confounded by missing factors from the
right-hand-side of the equation. That is, as briefly discussed in Chapter 4, if
there are missing factors that account for different systematic risks and
associated returns of assets, the $$\alpha _0$$ term will absorb them.
As such, $$\alpha _0$$ cannot be strictly interpreted to be an estimate
of the risk premium on the zero-beta portfolio—it can be biased up by
missing factor returns. Indeed, consistent with this logic and previously
discussed in Chapter 4, Jensen (1968) argued that $$\alpha _0$$ can
be used to test whether the market factor in the CAPM is sufficient to
explain excess asset returns. If missing factors exist, so-called Jensen’s
$$\alpha$$ will be greater than zero. In forthcoming chapters, we will
see that, based on real-world stock return data, many researchers have
found that the market model can be augmented with a number of other
factors that prove to have factor loadings that are significant in cross-
sectional tests.
Interestingly, the confounding effect of missing factors in the market
model does not necessarily contaminate the $$\lambda _0$$ intercept
estimate in cross-sectional regression (5.6). If there are missing factor
loadings in time-cross-sectional regression (5.6), this outcome will be
evidenced by low goodness-of-fit as measured by estimated $$R^2$$
values, rather than upward bias in $$\lambda _0$$. For this reason,
we favor cross-sectional tests of $$\lambda _0$$ as more reliable
indicators of the existence of the risk premium on the zero-beta portfolio.
Even so, the cross-sectional regression (5.6) is not without its own
problems. Early researchers found that measurement errors exist in the
estimated betas of individual stocks.6 To reduce this so-called errors-in-the-
variables problem, they recommended the use of portfolios in the
estimation of betas. This recommendation led researchers to sort stocks in
the U.S. and other markets into various portfolios to form test assets in
empirical tests of different asset pricing models.
Fama and French (2004) have provided an excellent review of the cross-
sectional regression evidence in the literature. As discussed there, in
addition to our review of studies in Chapter 4, early evidence by
Douglas (1969), Black et al. (1972), Miller and Scholes (1972), Blume and
Friend (1973), Fama and MacBeth (1973), and Fama and French (1992)
typically found that the slope $$\lambda _m$$ of the cross-sectional
relationship between excess portfolio returns and betas is lower than
predicted by the CAPM with estimated cross-sectional intercept
$$\lambda _0&gt;0$$. In their 2004 study, Fama and French repeated
these tests using U.S. stock market returns from the longer sample period
1928 to 2003. The results confirmed those of earlier studies. Together, these
studies’ results lend some degree of support for Black’s zero-beta
CAPM assumption that investors borrow at a rate higher than the riskless
rate in the financial market.
 Lastly, both the CAPM and the zero-beta CAPM assume that the market
portfolio is mean-variance efficient per Markowitz (1959). Thus, in
empirical tests of these models, as Jensen recognized, there is a joint
hypothesis: (1) the model is correct; and (2) the market portfolio is efficient.
In cross-sectional tests of k other factors that can be added to augment the
market factor in the market model, the $$\lambda _k$$s should equal
zero for the CAPM and zero-beta CAPM to hold. If these cross-sectional
regression coefficients are nonzero, then not only are these models rejected
but the proxy used to measure market portfolio returns is not efficient.
Hence, tests of $$\lambda _k$$s are akin to Jensen’s $$\alpha$$
in the sense that they test the validity of the CAPM and zero-beta
CAPM and their underlying efficient market portfolio assumption.
While beyond the scope of the present book’s introductory treatment of
asset pricing, sophisticated econometric tests of whether all the
$$\alpha _0$$ estimates for different portfolios (in time-series
regression specifications of asset pricing models) jointly equal zero have
been developed over time. As reviewed in more detail in Sect. 5.4, Gibbons
et al. (1989) developed an F-statistic of CAPM intercepts that also tests
whether the stock market index used to proxy the market portfolio is the
tangency portfolio.7 Of course, if this proxy is not efficient, according to the
Roll (1977) critique mentioned in Chapter 4, the CAPM cannot be tested.

5.3 Connection to Later Research

As a prelude to subsequent asset pricing research to be covered in
forthcoming chapters, Black’s two-factor zero-beta CAPM was the seminal
multifactor model. A multifactor model supplements the market factor with
one or more other factors. Extending the BJS findings, a landmark study by
Fama and French (1992) updated earlier evidence on the relationship
between average returns and CAPM beta for U.S. stocks and found that it is
even flatter than previously believed. Indeed, they found little or no
significant relationship to support the CAPM. Moreover, they found that
firm characteristics, such as size (measured by the market capitalization of
stocks equal to price times outstanding shares) and value (measured by the
ratio of book equity to market equity denoted BM) were significantly
related to average stock returns.
 Fama and French (2004) later interpreted these findings to mean that the
CAPM excludes important risks that investors use to price assets in
financial markets. They cited Merton’s (1973) intertemporal CAPM
(ICAPM) as one possible explanation, which we discuss next in Chapter 6.
In brief, ICAPM investors not only are concerned about the end-of-period
value of their portfolios of assets (i.e., wealth) but the ability to consume or
invest their wealth. Merton introduced state variables, including labor
income, goods prices, and portfolio opportunities, that investors use to
evaluate multiple risks. According to Fama and French (1992, p. 38),
investors seek efficient portfolios that are multifactor efficient earning the
highest expected returns per unit of multiple risks associated with the state
variables (or factors). The joint hypothesis is changed in this context to take
into account a multifactor model with multiple beta risks and a multifactor
efficient market portfolio. While Fama and French did not mention the
zero-beta CAPM as justification for multifactor models, it seems logically
relevant to their work, especially given that updated BJS evidence
motivated Black’s model.
Based on their 1992 findings, Fama and French (1993, 1995) proposed a
so-called three-factor model that is comprised of market, size, and value
factors. They documented evidence that beta coefficients associated with
the size and value factors are significantly related to average stock returns
but not beta related to the market factor. Regarding Merton’s ICAPM, they
argued that the size and value factors are not ICAPM state
variables themselves but are proxies for hidden (underlying) state variables.
As later noted in Fama and French (2004), due to this indirect connection to
state variables, their three-factor model was considered to be more
consistent with the arbitrage pricing theory (APT) of Ross (1976). This
theoretical model proposed that expected returns are a linear function of
multiple risk factors and their associated beta coefficients. If prices diverge
from the risk-adjusted price based on multiple risk factors, arbitrage by
investors will buy and sell assets until the linear relation between returns
and risks is established again.
The Fama and French research results on the significance of firm-level
characteristics, such as size and value, in better explaining average stock
returns were serious challenges to the CAPM and zero-beta CAPM. In
effect, these authors take the position that the $$\alpha _0$$ term in
time-series regressions is nonzero due to missing size and value factors.
 Deviations from asset pricing models, such as a nonzero
$$\alpha _0$$ term, that are not explained by risk factors are
considered to be market anomalies. The size and value anomalies are not
explained by CAPM beta risk. According to the efficient market hypothesis
(EMH) of Fama (1970), which led to his 2013 Nobel Prize in Economics,
asset prices should reflect all available information. In this world, asset
pricing models should have zero $$\alpha _0$$s after appropriate risk
adjustment. If $$\alpha _0$$ is nonzero, we can infer that an
anomaly exists that bears explanation by a missing systematic risk factor.
However, what if investors are not always rational? Now deviations
from asset pricing models could arise due to irrational investor
behavior rather than an unexplained risk not yet discovered. This is where
behavioralists enter the asset pricing picture. Behavioralists argue that
anomalies arise due to psychological human reasons. Irrational investor
responses due to overreaction, overconfidence, underreaction,
overextrapolation of past performance, etc., can cause asset prices to
fluctuate in ways not explained by the risk factors in asset pricing models
and therefore the EMH. Good examples of behavioral research papers are
by Debondt and Thaler (1985, 1987), Lakonishok et al. (1994), Barberis et
al. (1998), Daniel et al. (1998), Thaler (1999, 2016), and citations therein.
Do asset pricing models misprice assets due to bad model problems—
that is, not taking into account all the risk factors used by investors—or
does this happen due to behavioral issues related to human psychology?
This debate is certainly an interesting one! Continued research is needed to
more fully understand why asset mispricing occurs. Perhaps both EMH and
behavioral explanations are possible. In the long run after all information is
revealed, the EMH should be operative. But in the short run there may be
bursts of market behavior that are driven by short term, transitory risk
factors, and human behavior.
As an example, on October 19, 1987, which is known as Black Monday,
the Dow Jones Industrial Average (DJIA) experienced its largest one day
loss in the history of the U.S. stock market. World stock markets
plummeted also. Apparently, surging sell orders overwhelmed buy orders in
stock exchanges, and this order imbalance triggered computerized trading
as well as panic selling among investors. Weeks later the market more or
less rebounded from the sudden downturn. While transitory structural
aspects of the stock market played a central role in the collapse, market
behavior due to panic no doubt was responsible also. Black Monday was a
dramatic reminder of the short-run impact of temporary market risks as well
as irrational market behavior on asset prices.

5.4 GRS Test for Portfolio Efficiency

An important theoretical implication of the CAPM is that the market
portfolio M is mean-variance efficient. Unfortunately, the true market
portfolio is unobservable. For this reason, empirical tests of the CAPM are
forced to use proxies (like the value-weighted CRSP index) for the market
portfolio denoted m. Recognizing this shortcoming, Gibbons, Ross, and
Shanken (GRS) (1989) developed a time-series regression approach to test
whether any particular portfolio is ex ante mean-variance efficient (see
Cochrane (2005) for an excellent discussion of the GRS test).
The GRS test utilizes the following time-series regressions of asset
returns on a portfolio, or
$$\begin{aligned} r_{it} = \alpha _{i} + \beta _{i} r_{pt} + e_{it},
(5.7)
\end{aligned}$$
in which $$r_{it} = R_{it} - R_{ft}$$ is the excess return of asset
$$i = 1, \ldots , N$$ at time t, $$r_{pt} = R_{pt} - R_{ft}$$ is
the excess return on the portfolio whose efficiency is being tested, and
$$e_{it}$$ is the random disturbance term for asset i in period t. The
disturbances are assumed to be jointly normally distributed in each period
with mean zero and covariance matrix $$\Sigma$$ conditional on the
excess returns for portfolio p. It is assumed that the $$N\times N$$
covariance matrix is non-singular and returns are independent over time.
From general mean-variance portfolio mathematics, it follows that the
first-order condition for portfolio p to be mean-variance efficient (i.e., that
the portfolio is on the efficient frontier spanned by assets
$$i = 1, \ldots , N$$) is
$$\begin{aligned} E(r_{it}) = \beta _{im} E(r_{pt})
(5.8)
\end{aligned}$$
for all $$i = 1, \ldots , N$$. Thus, testing statistically the mean-
variance efficiency of portfolio p reduces to testing jointly the null
hypothesis
 $$\begin{aligned} H_0: \alpha _{i} = 0, \text { for all } i = 1, \ldots , (5.9)
N. \end{aligned}$$
A standard statistical test exists for simultaneously testing whether all
observations’ intercepts equal zero conditional on the excess return of
portfolio p. The test is a generalization of the usual t-test, known as the
Hotelling’s $$T^2$$-test, and therefore is distinct from the single
coefficient Student’s t-test. For this particular case, in which
$$r_{pt}$$ is a common independent variable in all regressions, the
test statistic for testing null hypothesis (5.9) becomes
$$\begin{aligned} F = \frac{T - N - 1}{N}\hat{\alpha
}_p'\hat{\Sigma }^{-1}\hat{\alpha }_p / (1 + \hat{\theta }_p^2), (5.10)
\end{aligned}$$
where
$$\hat{\alpha }_p' = (\hat{\alpha }_{1p}, \ldots , \hat{\alpha
}_{Np})$$
is the vector of OLS estimated $$\alpha _{ip}$$ intercepts (the prime
means vector transposition), $$\hat{\Sigma }^{-1}$$ is the inverse of
the OLS residual covariance matrix, and
$$\hat{\theta }_p = \bar{r}_p / s_p$$ is the Sharpe (1966) ratio of the
portfolio p with $$\bar{r}_p$$ the sample mean of the excess returns
of the portfolio p and $$s_p$$ the standard deviation of the portfolio
p. Under the null hypothesis, the F-test in Eq. (5.10) has an F-distribution
with degrees of freedom N and $$T - N - 1$$.
Gibbons, Ross, and Shanken modified the Hotelling $$T^2$$-test
to take into account whether portfolio p lies on the efficient frontier. They
showed that
$$\hat{\alpha }_p'\hat{\Sigma }^{-1}\hat{\alpha }_p = \hat{\theta
}^{*2} - \hat{\theta }_p^2$$
, where $$\hat{\theta }^{*2}$$ is the square of the ex post maximum
Sharpe ratio for the sample mean-variance efficient frontier portfolio
computed using all the N assets, and $$\hat{\theta }_p^2$$ is the
square of the Sharpe ratio of portfolio p defined above. As a result, the F-
statistic in Eq. (5.10) can be written as
(5.11)
 $$\begin{aligned} F = \frac{T - N - 1}{N}\frac{\hat{\theta
}_p^{*2} - \hat{\theta }_p^2}{1 + \hat{\theta }_p^2},
\end{aligned}$$
which gives a clear intuitive interpretation of the statistic. More specifically,
it indicates whether the observed deviation of the portfolio p from the
efficient frontier (which is the line through the tangent point of the mean-
variance frontier) can be inferred to be random noise. Also, it is notable that
$$\hat{\theta }^{*2} - \hat{\theta }_p^2 \ge 0$$ because portfolio p is
included in the efficient frontier calculations.8 We have left as an exercise
for the student to show that the F-statistic in Eq. (5.11) can be equivalently
presented as
$$\begin{aligned} F = \frac{T - N - 1}{N}\left( \left[ \frac{\sqrt{1
+ \hat{\theta }^{*2}}}{\sqrt{1 + \hat{\theta }_p^2}}\right] ^2 - (5.12)
1\right) , \end{aligned}$$
which is related to the form in Equation (7) of Gibbons et al. (1989, p.
1126).9
Example. To illustrate GRS testing, consider $$N = 48$$ industry
portfolios are available on Kenneth French’s data library10 and S &P500
index returns from finance.yahoo.com. R-code for downloading needed
data and computing the statistics is provided in Appendix B.

Fig. 5.2 Industry portfolios and S &P 500 index returns with ex post efficient frontier and the
market line

Using $$T = 261$$ monthly returns from January 2000 to

September 2021, Figure 5.2 shows a scatter plot of average monthly returns
and standard deviations of returns. Also, the ex post efficient frontier,
market line, tangency portfolio (green dot), and S &P 500 (red dot) are
shown in the graphs. We immediately observe that the proxy market
portfolio (S &P 500) is well below the efficient frontier, including the
tangency portfolio. In testing the mean-variance efficiency hypothesis in
relation (5.9), we can compute from the returns
$$\hat{\theta }_p = (\bar{r}_{\text {sp500}} - r_f) / s_ {\text
{sp500}} = (0.51 - 0.13) / 4.34 \doteq 0.088$$
. Using the fact that the Sharpe ratio of the tangency portfolio can be
computed as
$$\hat{\theta }^* = \sqrt{\bar{r}' \hat{\Omega }^{-1}\bar{r}}$$,
where $$\bar{r}$$ is the $$N+1$$ vector of averages of the N
excess stock returns and excess return of the S &P 500 index, and
$$\hat{\Omega }^{-1}$$ is the inverse of the sample covariance
matrix of these return series, we get $$\hat{\theta }^* = 0.782$$.
Thus, the f-statistic is:
$$F = \left( \frac{261 - 48 - 1}{48}\right) \left( \frac{0.782^2 - 0.088^2}
{1 + 0.088^2}\right) \doteq 2.65.$$
where $$f_1 = 212$$ numerator degrees of freedom and
$$f_2 = 48$$ denominator degrees of freedom, such that the p-value
is zero to at least three decimal places. We infer that the null hypothesis in
Eq. (5.9) is clearly rejected, which means that the S &P500 index is not an
efficient frontier portfolio.

5.5 Summary
Early stock return evidence using the market model to estimate the CAPM
found that the slope of the Security Market Line (SML) was flatter than
expected and the intercept term was higher than the predicted riskless
rate of interest (proxied by the government bond rate). Given this evidence,
Black (1972) proposed the zero-beta CAPM in which both long and short
positions in risky assets were allowed and investor borrowing could occur
at rates higher than the riskless rate. His asset pricing model led to a two-
factor model comprised of an efficient portfolio factor and zero-beta
portfolio factor. The efficient portfolio and zero-beta portfolio are assumed
to be uncorrelated (or orthogonal) to one another. Efficient portfolios are
changed from those located along a single line in the CAPM (extending
from the riskless rate to the tangent market portfolio on the
Markowitz efficient frontier of risky assets) to two possible cases: (1)
portfolios along the ray from the riskless rate to the tangent point of the
efficient frontier; and (2) a segment of the efficient frontier extending from
this tangent point to the end of the efficient frontier. The latter result is due
to the fact that investors are not allowed to borrow at the riskless rate. The
resultant zero-beta CAPM is more general than the CAPM because it holds
for any two portfolios located on the efficient frontier of the mean-variance
parabola and its zero-beta counterpart on the lower, inefficient boundary of
the parabola.
Empirical tests of the zero-beta CAPM are based on both time-series
regressions and cross-sectional regressions. The Fama and MacBeth (1973)
two-step procedure combines both of these regressions in an effort to
estimate the relation between betas and future returns of asset portfolios.
The time-series regression of the market model for the CAPM yields an
intercept known as Jensen’s $$\alpha$$ which should be zero for all
assets. If they are nonzero, the implication is that the zero-beta CAPM is
valid; however, other missing factors could explain nonzero
$$\alpha$$s also. The cross-sectional regression yields an intercept
denoted $$\lambda _0$$ that measures the risk premium on the zero-
beta portfolio. If $$\lambda _0&gt;0$$, then the zero-beta CAPM’s
assumption that investors borrow at a rate greater than the riskless rate is
supported. Also, if other risk factors than the market factor are significant in
cross-sectional tests, then the zero-beta CAPM does not hold.
Research by Fama and French (1992, 1993, 1995) showed that size and
value are significant firm-level characteristics that help to explain average
stock returns. Thus, they rejected the zero-beta CAPM in addition to its
precursor the CAPM. Because any test of the CAPM or zero-beta CAPM is
a joint test that the model is valid and the portfolio proxy is an efficient
portfolio, their findings suggested that efficient portfolios must be
multifactor efficient in terms of multiple risk factors. By way of
explanation, they cited the ICAPM of Merton (1973) and APT of
Ross (1976) as theoretical models that are consistent with their three-factor
model containing market, size, and value factors.
While the efficient market hypothesis (EMH) suggests that prices reflect
all available information to evaluate risks in financial markets,
behavioralists argue that mispricing can happen due to irrational actions of
investors affected by human psychology. If behavioralists are right, then
asset pricing faces serious challenges. It is likely that short-run market
anomalies can occur due to irrational investor behavior as well as short run,
transitory market risk factors, but long-run anomalies are attributable to
unknown risk factors. Regarding the latter, if investors are rational (for the
most part), asset pricing models should continue to evolve and increasingly
explain the average returns of stocks (and other asset classes such as bonds,
commodities, real estate, etc.) in the long run. As we will see in upcoming
chapters, researchers are making progress!
In his assessment of the zero-beta CAPM, Sharpe (1973) critically
commented that, given any efficient portfolio, its zero-beta portfolio is not
directly identifiable. In asset pricing models, researchers typically employ
proxies for the market portfolio and the riskless rate—namely, the value-
weighted index of all stocks in the market and the government bond rate.
However, it is not clear how to proxy the zero-beta portfolio. With no such
ready proxy at hand, researchers typically do not include a zero-beta
portfolio rate of return in their specifications of different models. Instead, a
riskless rate proxy is almost always employed as the borrowing rate in
empirical models. Circumventing the zero-beta portfolio identification
problem, Chapter 10 reviews a new form the zero-beta CAPM dubbed the
ZCAPM. The ZCAPM enables estimation of the zero-beta CAPM with
readily available stock return data that does not require a proxy for the zero-
beta portfolio rate.

Questions
1.
Black (1972) developed the zero-beta CAPM due to some unrealistic
assumptions of the CAPM. What assumptions did he relax?
2.
Black et al. (1972) found that the market model needed to be modified
to better fit U.S. stock returns. What did they propose as a better fitting
model?
3.
Black (1972) proposed a zero-beta CAPM without and with a riskless
rate. Write the equations for these two models. Can investors borrow at
the riskless rate in the second model?
4. In the CAPM, the Security Market Line (SML) defines combinations of
the riskless asset and market portfolio M that investors can hold. By
contrast, under Black’s zero-beta CAPM, what are two ways for
investors to hold efficient portfolios? Draw a diagram to illustrate your
answer. Where is the market portfolio M in this diagram? How can the
investor hold M. Assuming the investor holds an efficient portfolio,
using your diagram, how can the uncorrelated zero-beta portfolio
counterpart be located?
5.
Assume that the investor holds risky efficient portfolio I that is a
combination of market portfolio M and its uncorrelated zero-beta
portfolio counterpart ZI. How can we write the zero-beta CAPM using
these two risky portfolios?
6.
Assume that you run a time-series regression of the market model.
Next, you run a cross-sectional regression using one-month-ahead
returns. In the cross-sectional regression, how should you interpret the
estimated market prices of beta risk, or $$\lambda _m$$, and the
intercept term, or $$\lambda _0$$? How is $$\lambda _m$$
related to the expected market risk premium in the CAPM?
7.
Fama and French published a series of studies in the 1990s that are
important to the validity of the CAPM and zero-beta CAPM. What did
they find in these studies? What did they propose as an alternative asset
pricing model?
8.
What is a market anomaly? What might explain market anomalies?
Give two explanations. What does the efficient market hypothesis
(EMH) have to say about market anomalies?
9.
Can financial markets be irrational? Give an example from stock
market history.

Problems
1.
How can the zero-beta CAPM be empirically tested with real-world
data? Discuss two tests using equations based on the market model.
Which test is more reliable and why?
2. Show that the GRS statistic
$$F = \frac{T - N - 1}{N} \frac{\hat{\theta }^{*2} - \hat{\theta
}_p^2}{1 + \hat{\theta }_p^2}$$
can be equivalently written as
 $$F = \frac{T-N-1}{N}\left(\left[ \frac{\sqrt{1 + \hat{\theta
}^{*2}}}{\sqrt{1 + \hat{\theta }_p^2}}\right] ^2 - 1\right).$$
3.
Using 12 industry portfolios from Kenneth French’s data library,
replicate the GRS testing for mean-variance efficiency of the market
portfolio, available in
FF_Research_Data_Factors_CSV. Use monthly data for the time
period January 2000 to September 2021. Verify that the test statistics in
Eqs. (5.10) and (5.11) produce the same results.

Appendix A: Mathematical Derivation of the

Zero-Beta CAPM
Here we provide a mathematical proof of Black’s zero-beta CAPM. The
derivation relies heavily on the work of Kolari et al. (Chapter 2, 2021). To
begin, according to Markowitz (1959), the minimum variance of an
efficient portfolio I of assets with weights that sum to one and expected
return $$E({R}_I)$$ can be defined as follows:
$$\begin{aligned} \underset{\omega }{\text {Min}}\frac{1}
{2}\omega '\Sigma \omega +\eta [E({R}_I)-\omega 'E({R})]+\kappa (A5.1)
(1-\omega 'e), \end{aligned}$$
where $$\omega$$ is a vector of weights (with dimension
$$n\times 1$$), e is a vector of ones (with dimension $$n\times 1$$),
$$E({R}) = [E({R}_1), E({R}_2), \cdots , E({R}_n)]'$$ is the expected
one-period return vector for n assets (with dimension $$n\times 1$$), and
$$\Sigma$$ is the covariance matrix for n assets (with dimension
$$n\times n$$). First-order conditions with respect to $$\omega$$,
$$\eta$$, and $$\kappa$$ are:
$$\begin{aligned} \Sigma \omega -\eta E({R})-\kappa e= &amp;
{} 0\nonumber \\ E({R}_I)-\omega 'E({R})= &amp; {} 0\nonumber (A5.2)
\\ 1-\omega ' e= &amp; {} 0. \end{aligned}$$
Upon solving to get the optimal weights $$\omega ^*$$, we obtain:
(A5.3)
 $$\begin{aligned} \omega ^*= &amp; {} \frac{CE({R}_I)-A}
{BC-A^2}\Sigma ^{-1}E({R}) +\frac{B-\alpha E({R}_I)}{BC-
A^2}\Sigma ^{-1}e\nonumber \\\equiv &amp; {} \phi +\psi
E({R}_I), \end{aligned}$$
where
$$\begin{aligned} A\equiv &amp; {} E({R})'\Sigma
^{-1}e\nonumber \\ B\equiv &amp; {} E({R})'\Sigma
(A5.4)
^{-1}E({R})\nonumber \\ C\equiv &amp; {} e'\Sigma ^{-1}e
\end{aligned}$$
are known scalars, and
$$\begin{aligned} \phi\equiv &amp; {} \frac{B\Sigma ^{-1}e-
A\Sigma ^{-1}E({R})}{BC-A^2}\nonumber \\ \psi\equiv &amp; {}
(A5.5)
\frac{C\Sigma ^{-1}E({R})-A\Sigma ^{-1}e}{BC-A^2}
\end{aligned}$$
are known $$n\times 1$$ vectors. Based on these equations, the variance
of efficient portfolio I can be written as:
$$\begin{aligned} \sigma ^2_{I}=\frac{1}{C} + \frac{C\left[
(A5.6)
E({R}_{I})-\frac{A}{C}\right] ^2}{BC-A^2}. \end{aligned}$$
In line with the CAPM, we can write the equation of the line through the
tangent point $$[E({R}_M),\sigma _M]$$ (i.e., market portfolio M) as:
$$\begin{aligned} E({R}_{ZI})-E({R}_M)=\frac{BC-A^2}
{C[E({R}_M)-\frac{A}{C}]}\sigma _M(\sigma _{ZI}-\sigma (A5.7)
_{M}), \end{aligned}$$
where $$\frac{BC-A^2}{C[E({R}_M)-\frac{A}{C}]}\sigma _M$$ is the
slope of the line as defined by Merton (1972, p. 1856). By replacing
$$E({R}_{ZI})$$ with the riskless rate $$R_f$$, we can achieve the
highest (most efficient) slope of the Capital Market Line (CML) (i.e.,
$$\sigma _{ZI} = 0$$). Subsequently, we obtain the familiar slope of the
CML defined as $$[E({R}_M)-R_f]/\sigma ({R}_M)$$.
In Black’s zero-beta CAPM, the expected return on any risky asset
portfolio P is a function of the expected returns on efficient portfolio I and
its zero-beta portfolio counterpart ZI. As proof, we begin by defining the
covariance between the returns of portfolios I and P as follows:
 $$\begin{aligned} Cov ({R}_{I},{R}_{P})=\omega ^{M'}\Sigma (A5.8)
\omega ^P=[\phi +\psi E({R}_I)]'\Sigma \omega ^P. \end{aligned}$$
In terms of A, B, and C, this covariance term can be specified as:
$$\begin{aligned}= &amp; {} \left( \frac{B\Sigma ^{-1}e-
A\Sigma ^{-1}E({R})}{BC-A^2}+\frac{C\Sigma ^{-1}E({R})-
A\Sigma ^{-1}e}{BC-A^2}E({R}_I)\right) '\Sigma \omega
^P\nonumber \\= &amp; {} \frac{Be'\Sigma ^{-1}\Sigma \omega ^P-
AE({R})'\Sigma ^{-1}\Sigma \omega ^P}{BC- (A5.9)
A^2}+\frac{CE({R}_I)E({R})'\Sigma ^{-1}\Sigma \omega ^P-
AE({R}_I)e'\Sigma ^{-1}\Sigma \omega ^P}{BC-A^2}\nonumber
\\= &amp; {} \frac{B-AE({R}_P)+CE({R}_I)E({R}_P)-AE({R}_I)}
{BC-A^2}. \end{aligned}$$
Solving for $$E({R}_P)$$ yields:
$$\begin{aligned} E({R}_P)= &amp; {} \frac{AE({R}_I)-B}
{CE({R}_I)-A}+ Cov ({R}_{I},{R}_{P})\left( \frac{BC-A^2}
{CE({R}_I)-A}\right) \nonumber \\= &amp; {} \frac{A}{C}-
(A5.10)
\frac{BC-A^2}{C^2E({R}_I)-\frac{A}{C}}+ Cov ({R}_{I},
{R}_{P})\frac{\sigma ^2_{I}}{\sigma ^2_{I}} \left( \frac{BC-
A^2}{CE({R}_I)-A}\right) , \end{aligned}$$
where
$$\sigma ^2_{I}=\frac{1}{C}+\frac{C[E({R}_I)-\frac{A}{C}]^2}
{BC-A^2}$$
. Expanding the above equation, we get:
$$\begin{aligned} E({R}_P)=\frac{A}{C}-\frac{BC-A^2}
{C^2[E({R}_I)-\frac{A}{C}]} +\frac{ Cov ({R}_{I},{R}_P)}
{\sigma ^2_{I}}\left( E({R}_I)-\frac{A}{C}+\frac{BC-A^2} (A5.11)
{C^2[E({R}_I)-\frac{A}{C}]}\right) .\nonumber \\
\end{aligned}$$
The intercept term in this equation corresponding to portfolio return
$$\frac{A}{C}-\frac{BC-A^2}{C^2[E({R}_I)-\frac{A}{C}]}$$ is
orthogonal to efficient portfolio return $$E({R}_I)$$. As such, it
represents the expected return on the zero-beta portfolio or
$$E({R}_{ZI})$$. To see this, note that:
(A5.12)
 $$\begin{aligned} \omega ^{ZI'}\Sigma \omega ^I=[\phi +\psi
E({R}_{ZI})]'\Sigma [\phi +\psi E({R}_I)]=0\end{aligned}$$

$$\begin{aligned} B-AE({R}_{ZI})+CE({R}_I)E({R}_{ZI})-
(A5.13)
AE({R}_I)=0\end{aligned}$$

$$\begin{aligned} E({R}_{ZI})=\frac{AE({R}_I)-B}
{CE({R}_{I})-A}=\frac{A}{C}-\frac{BC-A^2}{C^2[E({R}_I)- (A5.14)
\frac{A}{C}]}. \end{aligned}$$
Finally, upon re-writing Eq. (A5.11), Black’s zero-beta CAPM is obtained:
$$\begin{aligned} E({R}_{P})=E({R}_{ZI})+\beta
(A5.15)
_P[E({R}_I)-E({R}_{ZI})]. \end{aligned}$$
Appendix B: R Snippet for the GRS Testing
Example in Section 5.4

References
Barberis, M., A. Shleifer, and R. Vishny. 1998. A model of investor sentiment. Journal of Financial
Economics 49: 307–343.
[Crossref]

Black, F. 1972. Capital market equilibrium with restricted borrowing. Journal of Business 45: 444–
454.
[Crossref]

Black, F., M.C. Jensen, and M. Scholes. 1972. The capital asset pricing model: Some empirical tests.
In Studies in the Theory of Capital Markets, ed. M.C. Jensen. New York, NY: Praeger.

Blume, M. 1970. Portfolio theory: A step towards its practical application. Journal of Business 43:
152–174.
[Crossref]

Blume, M., and I. Friend. 1973. A new look at the capital asset pricing model. Journal of Finance
28: 19–33.
[Crossref]

Cochrane, J.H. 2005. Asset Pricing. Revised. Princeton, NJ: Princeton University Press.

Daniel, K., D. Hirshleifer, and A. Subrahmanyam. 1998. Investor psychology and security market
under- and overreactions. Journal of Finance 53: 1839–1885.
[Crossref]

De Bondt, W.F.M., and R. Thaler. 1985. Does the stock market overreact? Journal of Finance 40:
793–805.
[Crossref]

De Bondt, W.F.M., and R. Thaler. 1987. Further evidence of stock market overreaction and
seasonality. Journal of Finance 42: 557–581.
[Crossref]

Douglas, G.W. 1969. Risk in the equity markets: An empirical appraisal of market efficiency. Yale
Economic Essays 9: 3–45.

Fama, E.F. 1968. Risk, return, and equilibrium: Some clarifying comments. Journal of Finance 23:
29–40.
[Crossref]

Fama, E.F. 1970. Efficient capital markets: A review of theory and empirical work. Journal of
Finance 25: 383–417.
[Crossref]

Fama, E.F., and K.R. French. 1992. The cross-section of expected stock returns. Journal of Finance
47: 427–465.
[Crossref]

Fama, E.F., and K.R. French. 1993. The cross-section of expected returns. Journal of Financial
Economics 33: 3–56.
[Crossref]

Fama, E.F., and K.R. French. 1995. Size and book-to-market factors in earnings and returns. Journal
of Finance 50: 131–156.
[Crossref]

Fama, E.F., and K.R. French. 2004. The capital asset pricing model: Theory and evidence. Journal of
Economic Perspectives 18: 25–46.
[Crossref]

Fama, E.F., and J.D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy 81: 607–636.
[Crossref]

Ferson, W.E. 2019. Empirical Asset Pricing: Models and Methods. Cambridge, MA: The MIT Press.

Ferson, W., A.F. Siegel, and J.L. Wang. 2019. Asymptotic distributions of tests of portfolio efficiency
and factor model comparisons with conditioning information. Working paper, University of Southern
California and University of Washington.

Friend, I., and M. Blume. 1970. Measurement of portfolio performance under uncertainty. American
Economic Review 60: 607–636.

Gibbons, M.R., S.A. Ross, and J. Shanken. 1989. A test of the efficiency of a given portfolio.
Econometrica 57: 1121–1152.
[Crossref]

Green, R.C., and B. Hollifield. 1992. When will mean-variance efficient portfolios be well
diversified? Journal of Finance 47: 1785–1809.
[Crossref]

Jagannathan, R., and T. Ma. 2003. Risk reduction in large portfolios: A role for portfolio weight
constraints. Journal of Finance 58: 1651–1684.
[Crossref]

Jensen, M.C. 1968. The performance of mutual funds in the period 1945–1964. Journal of Finance
23: 389–416.
[Crossref]

Kallberg, J.G., and W.T. Ziemba. 1983. Comparison of alternative utility functions in portfolio
selection problems. Management Science 29: 1257–1276.
[Crossref]

Kroll, Y., H. Levy, and H. Markowitz. 1984. Mean-variance versus direct utility maximization.
Journal of Finance 39: 47–61.
[Crossref]

Lakonishok, J., A. Shleifer, and R.W. Vishny. 1994. Contrarian investment, extrapolation, and risk.
Journal of Finance 49: 1541–1578.
[Crossref]

Levy, H. 1983. The capital asset pricing model: Theory and empiricism. Economic Journal 93: 145–
165.
[Crossref]

Levy, M. 2007. Conditions for a CAPM equilibrium with positive prices. Journal of Economic
Theory 137: 404–415.
[Crossref]

Levy, M., and Y. Ritov. 2010. Mean-variance efficient portfolios with many assets: 50% short.
Quantitative Finance 10: 1–11.

Markowitz, H.M. 1959. Portfolio Selection: Efficient Diversification of Investments. New York, NY:
John Wiley & Sons.

Merton, R.C. 1972. An analytical derivation of the efficient portfolio frontier. Journal of Financial
and Quantitative Finance 7: 1851–1872.
[Crossref]

Merton, R.C. 1973. An intertemporal capital asset pricing model. Econometrica 41: 867–887.
[Crossref]

Miller, M., and M. Scholes. 1972. Rates of return in relation to risk: A re-examination of some recent
findings. In Studies in the Theory of Capital Markets, ed. M.C. Jensen, 47–78. New York, NY:
Praeger.

Pulley, L.B. 1981. General mean-variance approximation to expected utility for short holding
periods. Journal of Financial and Quantitative Analysis 16: 361–373.
[Crossref]

Roll, R. 1977. A critique of the asset pricing theory’s tests, part I: On past and potential future
testability of the theory. Journal of Financial Economics 4: 129–176.
[Crossref]

Roll, R. 1980. Orthogonal portfolios. Journal of Financial and Quantitative Analysis 15: 1005–1012.
[Crossref]

Ross, S.A. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13: 341–
360.
[Crossref]

Sharpe, W.F. 1966. Mutual fund performance, Journal of Business 39: 119–138.
[Crossref]

Sharpe, W.F. 1973. The capital asset pricing model: Traditional and “zero-beta” versions.
Presentation at the 1973 annual meetings of the Midwest Finance Association, 1–15.

Thaler, R.H. 1999. The end of behavioral finance. Financial Analysts Journal 55: 12–17.
[Crossref]

Thaler, R.H. 2016. Behavioral economics: Past, present, and future. American Economic Review 106:
1577–1600.
[Crossref]

Footnotes
1 This graph is based on Figure 4 in Sharpe’s (1973) conference presentation to the Midwest Finance
Association.

2 See equation (40) in Black (1972, p. 454).

3 See studies by Roll (1977), Levy (1983), and Levy (2007).

4 For example, see Pulley (1981), Kallberg and Ziemba (1983), Levy (1983), Kroll et al. (1984),
Green and Hollifield (1992), Jagannathan and Ma (2003), Levy and Ritov (2010), and others.

5 For example, assuming all betas for assets are positive, long and short positions would be needed
to get a portfolio beta equal to zero. Since it is a zero-investment portfolio like other factors, we use
the notation $$\lambda$$ to acknowledge this fact.

6 For example, see Blume (1970), Friend and Blume (1970), and Black et al. (1972).

7 See recent work by Ferson et al. (2019) for further discussion and citations of other studies that
have progressively refined these statistical tests.
8 In other words, when the riskless rate is available, $$\hat{\theta }^{*2}$$ is the square of the
slope of the tangent line from the origin to the (ex post) efficient frontier, whereas $$\theta _p^2$$
is the slope of the line from the origin to portfolio p in standard deviation and excess return space.

9 The intuition of representation (5.12) is as follows. Since a line in the standard deviation (
$$\sigma$$) mean excess return ( $$\mu$$) space through the origin is of the form
$$\mu = \theta \sigma$$, the square of the length of the line (by Pythagoras) is
$$\sigma ^2 + \theta ^2\sigma ^2 = (1 + \theta ^2)\sigma ^2$$. As a result, the ratio
$$\sqrt{1 + \hat{\theta }^{*2}} / \sqrt{1 + \hat{\theta }_p^2}$$ in Eq. (5.12) is the ratio of the
lengths at any $$\sigma &gt; 0$$ of the tangent line with slope $$\hat{\theta }^{*2}$$ and the
line of portfolio p with slope $$\hat{\theta }_p$$. Therefore the ratio
$$[(1 + \hat{\theta }^{*2})\sigma ^2] / [(1 + \hat{\theta }_p^2)\sigma ^2] = (1 + \hat{\theta
}^{*2}) / (1 + \hat{\theta }_p^2)$$
is the ratio of the squared lengths of the lines. The closer this ratio is to the lower bound of one, the
closer the portfolio p is to the efficient frontier.

10 https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_6

6. Alternative CAPM Specifications

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_6

Weak early empirical evidence with respect to the CAPM inspired a variety
of alternative forms proposed by researchers to take into account realistic
aspects of capital markets not considered in its original form. In the last
chapter we covered Black’s (1972) zero-beta CAPM allowing short sales
and borrowing at rates greater than the riskless rate. Here we review further
extensions of the CAPM, including the foundational and more general
intertemporal CAPM (ICAPM) which spawned a number of new models
such as the international asset pricing model (IAPM), consumption CAPM
(CCAPM), production CAPM (PCAPM), and conditional CAPM. The latter
conditional CAPM allows for time-varying risk parameters in any asset
pricing model to take into account the business cycle with economic
expansions and recessions. Over the years, different forms of the
aforementioned models have been developed by researchers. In general,
they attest to the widespread interest in the CAPM by academics and
professionals in finance.
Unfortunately, the aforementioned extensions of the ICAPM did not
consistently cure the weak empirical evidence problem. While they
continue to draw the interest of many researchers, most asset pricing
researchers nowadays emphasize multifactor models. The latter models
augment the market factor by adding many new factors proposed by
researchers. Multifactor models are considered to be related to the
intertemporal CAPM (ICAPM) by Merton (1973) covered in this chapter in
addition to the Arbitrage Pricing Model (APT) by Ross (1976) discussed in
forthcoming Chapter 7. In Chapters 8 and 9 we will see that multifactor
models do a better job of explaining stock returns than the CAPM and its
companion models presented in this chapter. However, in Chapter 10 we
return to the CAPM with an alternative form of the zero-beta CAPM
dubbed the ZCAPM by Kolari et al. (2021). Recent empirical evidence
discussed there shows that the ZCAPM outperforms popular multifactor
models in cross-sectional tests. Hence, the development of alternative
CAPMs continues to hold promise in asset pricing.

6.1 Intertemporal CAPM

The CAPM is considered to be a static model in the sense that it is based on
a single-period. Considered by everyone to be a tour de force,
Merton (1973) extended the CAPM to a multiperiod setting. Upon doing so,
an intertemporal CAPM (ICAPM) was derived that yielded new insights
into equilibrium asset prices. For example, even if no systematic market
risk exists, the expected return on risky assets can exceed the riskless rate.
Other factors related to state variables that shape future investment
opportunities can contribute to expected returns. In the CAPM, investment
opportunities are fixed and the market factor is the only risk factor.1 Upon
allowing the future opportunity set available to investors to change over
time in response to state-dependent variables, Merton showed that asset
pricing is more complicated than the CAPM would predict. Indeed, there
are multiple potential factors associated with different state variables that
can influence asset prices.
 Merton examined the portfolio selection behavior of investors who
maximize their lifetime consumption and trade continuously over time. He
assumed that the demand function for assets by investors is affected by the
possibility of uncertain changes in future investment opportunities.
Changing investment opportunities are state-dependent in the ICAPM—for
example, good and bad economic states related to times of prosperity versus
recession.
The ICAPM makes similar assumptions to the CAPM: limited liability
of assets; no transaction costs and taxes; no wealth constraints on the ability
to buy and sell assets at market prices; capital markets in equilibrium,
borrowing and lending at the same rate of interest, and short sales of assets
(which is allowed in Black’s [1972] zero-beta CAPM version of the
CAPM). However, he adds a crucial new assumption: continuous trading of
assets over time. This assumption allows investors to revise their portfolios
over time as they learn new market information. Even daily revisions could
be made by investors. Taking this logic one step further, Merton adopted a
continuous-time process of asset pricing in capital markets, wherein the
price process follows what is known as geometric Brownian motion.2 This
random motion implies a lognormal return distribution for asset returns.
The CAPM assumes homogeneous expectations among investors about the
normal joint distribution of asset returns that are unchanging over time.
However, a changing investment opportunity set implies that homogeneous
expectations break down. Now asset prices are untethered from a fixed
return distribution.
Given the risk of a changing investment opportunity set due to
underlying state variables, investors seek to hedge (or insure) against
“unfavorable” shifts in such opportunities. Resultant risk hedges employed
by investors give rise to new asset pricing factors unrelated to the CAPM
market factor. To demonstrate these concepts, he introduced a changing
riskless rate as a candidate for such a state variable. The riskless rate is
known to investors in the beginning of the coming period but is changing
stochastically (or randomly) thereafter over time. Due to a stochastic
riskless rate, investors are exposed to possible shifts in the investment
opportunity set over time. Assuming that the (stochastic) riskless rate is
sufficient to describe changes in the investment opportunity set, he
proposed a three-fund theorem with the following portfolios chosen by
investors in equilibrium: (1) the riskless asset (f), (2) market portfolio (M),
and (3) an asset or portfolio that is perfectly negatively correlated with the
riskless rate (N) (i.e., hedge for the riskless asset). An example of the latter
N asset could be a long-term bond with returns negatively correlated to
changes in short-term interest rates. Using these assumptions and capital
market conditions, Merton derived a continuous-time version of the single-
period CAPM or intertemporal CAPM (ICAPM):3
$$\begin{aligned} E({R}_i) = R_f + \beta _{i,M}[E({R}_M) - R_f]
(6.1)
+ \beta _{i,N}[E({R}_N) - R_f], \end{aligned}$$
where $${R}_i$$ is the instantaneous return of stock i at any instant
of time t, $$R_f$$ is the instantaneous riskless rate,
$${R}_M$$ is the instantaneous market return, and $${R}_N$$
is the instantaneous return on a portfolio that has perfect negative
correlation with the riskless rate. The risk parameters in the ICAPM are
defined as follows:
$$\begin{aligned} \beta _{i,M} = \frac{\sigma _i}{\sigma _M}\cdot
\frac{\rho _{iM} - \rho _{iN}\rho _{NM}}{1 - \rho _{NM}^2}, \quad
(6.2)
\beta _{i,N} = \frac{\sigma _i}{\sigma _N}\cdot \frac{\rho _{iN} -
\rho _{iM}\rho _{NM}}{1 - \rho _{NM}^2} \end{aligned}$$
with various correlations among the three portfolios in this expression of
the general form $$\rho _{kl} = \textit{Cor}(R_k, R_l)$$ for any two
assets k and l.
Equation (6.1) indicates that investors are not only compensated for
bearing market risk but also for bearing the risk of unfavorable shifts in the
investment opportunity set. If correlations $$\rho _{iN} = 0$$ for all
securities i, then it must be true that $$\rho _{NM} = 0$$, where M is
the market portfolio containing all securities. Also, because $$R_N$$
is perfectly negatively correlated with $$R_f$$ by definition, the
correlations of stocks with the future riskless rate $$R_f$$ would be
zero. In this special case, the ICAPM reduces to the traditional CAPM in
which $$\beta _{i,N} = 0$$ and
$$\beta _{i,M} = (\sigma _i / \sigma _M)\rho _{iM} = (\sigma _i
\sigma _M/ \sigma ^2_M)\rho _{iM} = Cov ({R}_i, {R}_M)/\sigma
^2_M$$
. Since security N has a zero-beta due to zero correlation with market
portfolio M (or $$\rho _{NM} = 0$$), it is tantamount to the zero-
beta portfolio of Black (1972). Notice also that, if there is no systematic
market beta risk such that $$\beta _{i,M} = 0 = \rho _{iM}$$, but
there is beta risk associated with security N so that
$$\beta _{i,N}&gt;0$$, the expected return on the ith asset will be
greater than the riskless rate, unlike the traditional CAPM. Of course, in the
likely case that $$\rho _{iN} &gt; 0$$, asset prices are determined by
all risk factors that affect changing future investment opportunities.
On a practical level, we can specify the ICAPM as a multifactor
empirical model. To simplify notation, we use excess returns over the
riskless rate denoted with a lowercase r. Following Merton, we include a
number of state variables that can affect the expected excess returns of the
ith asset:
$$\begin{aligned} r_{it}= \alpha _0 + \beta _{i,M} r_{Mt} + \beta
_{i,s_1} s_{1t} + \beta _{i,s_2} s_{2t} + \cdots + \beta _{i,s_k} (6.3)
s_{kt} +\epsilon _{it}, \end{aligned}$$
where there are k state variables denoted $$s_{jt}$$,
$$j = 1, \ldots , k$$. The multiple regression coefficients in empirical
model (6.3) are beta risk measures for each asset. Each beta, or
$$\beta _{i, s_j}$$, contains a covariance term
$$Cov (r_{it}, s_{jt})$$, $$j = 1, \ldots , k$$, that determines
its specific systematic risk. In place of the k state variables in this equation,
we equivalently could use the hedging portfolio returns for the risk
associated with each state variable.
Merton leaves open the selection of state variables or definition of
hedge portfolio returns in the ICAPM. His model identifies only the
necessary properties for variables to qualify as state variables. As we will
see, forthcoming chapters introduce various multifactors that augment the
market factor in asset pricing models. To help visualize Merton’s ICAPM,
consider a two-factor model with the market factor plus a state
variable denoted $$s_1$$. Now total systematic risk is comprised of
beta coefficients $$\beta _{i,M}$$ and $$\beta _{i,s_1}$$
associated with the market factor and the state variable factor, respectively.
Figure 6.1 provides a three-dimensional graph of this asset pricing model
that shows the relation between expected excess returns and the two beta
risks. Assets with relatively low $$\beta _{i,s_1}$$ risk are shown in
dark dots and those with higher $$\beta _{i,s_1}$$ risk in gray dots
(located further back in the three-dimensional space). As state variables are
added to the ICAPM, more return/risk dimensions are created. How many
dimensions of systematic risk do you think investors use in capital markets?

Fig. 6.1 A three-dimensional depiction of an empirical ICAPM with two systematic beta risk
factors: (1) market factor beta $$\beta _{i,M}$$ , and (2) state variable beta $$\beta _{i,s1}$$

At the end of his paper, reflecting on the ICAPM, Merton (1973, pp.
885–886) suggested future directions for research:

The model is robust in the sense that it can be extended in an

obvious way to include effects other than shifts in the investment
opportunity set. Two important factors not considered are wage
income and many consumption goods whose relative prices are
changing over time. In a more complete model the three-fund
theorem ... will generalize to an m-fund theory. Also, there was no
discussion of the supply side, given a micro theory of the firm ... to
close the model.

Fig. 6.2 Alternative CAPM specifications based on the intertemporal CAPM (ICAPM)

This forward-looking statement about the ICAPM turned out to be an

accurate prediction of future equilibrium asset pricing models. As we will
see in this chapter, new CAPMs were proposed by researchers that expand
the breadth of the investment opportunity set to international markets
(IAPM) as well as lifetime consumption behavior on the demand side of the
market (CCAPM) and firm production on the supply side (PCAPM). Also,
conditional CAPM methods were introduced to better capture time-varying
risk parameters in asset pricing models. All these models are based on
ICAPM as depicted in Fig. 6.2. The ICAPM provided a roadmap for CAPM
researchers.
6.2 International Asset Pricing Model (IAPM)
6.2.1 World Market Factor
Solnick (1974a, b) extended the CAPM to an international asset pricing
model (IAPM). The IAPM posits that equilibrium security prices are
determined in an integrated world market comprised of different national
markets. That is, capital flows move freely between different national stock
markets in an inter-connected global market. By contrast, the CAPM
posited a single national market, in which investors choose between the
market portfolio and the riskless asset known as the two fund theorem. In
Solnik’s world market model, there is more than one riskless asset due to
different interest rates in different countries. Also, there is exchange rate
risk due to changes in the market values of different national currencies
held by investors around the world. An international three-fund theorem is
proposed, wherein investors can hold: (1) a world portfolio of stocks with
no exchange rate risk (i.e., the market portfolio), (2) a portfolio of bonds
with exchange rate risk but no market risk, and (3) the riskless asset of their
own country. He derived the IAPM from the national market ICAPM of
Merton (1973). Based on Merton’s equilibrium pricing framework, Solnik
(1974a, p. 515) specified the IAPM as follows:
$$\begin{aligned} E({R}_{i})-R_f=\beta _i [E({R}_{W})-R_B)],
(6.4)
\end{aligned}$$
where $$R_f$$ is the country riskless rate, $$E({R}_{W})$$ is the
world stock market equity return (or market portfolio return), and
$$R_{B}$$ is the average interest rate in the world or international bond
rate. The riskless bond in a country is riskless to national investors but for
foreign investors it represents a pure exchange risk asset that allows them to
hedge their equity investment from exchange rate risk (i.e., a short position
in the foreign bond). The $$\beta _i$$ coefficient measures the
international systematic risk associated with the world stock market
portfolio W. Bond portfolios f and B are different from one another. Notice
that the world stock and bond portfolios are independent of investors’ risk
preferences and national citizenship. The risk premium of an asset in a
country over its national riskless rate is proportional to its international
systematic risk premium. Also, the IAPM says that all investors should buy
a single world equity fund with market value weights.
Solnik (1974b, p. 541) stated the market model form of the IAPM as
follows:
$$\begin{aligned} R_{it}-R_{ft}=\alpha _0 + \beta _i (R_{wt}-
(6.5)
R_{bt})+\epsilon _{it}, \end{aligned}$$
where $$R_{it}-R_{f}$$ is asset i’s excess return over the national
riskless rate at time t, $$R_{wt}-R_{bt}$$ is the international equity
market excess return over the average world bond rate using proxies for W
and B, and $$\epsilon _{it}$$ are independent and normally
distributed error terms. Solnick proceeded to test this international market
model using stock returns in 10 major industrialized countries with data
from March 1966 to April 1971. A value-weighted world stock market
index return was constructed, whose weights were used to compute an
average world bond index also. Regressing national excess market returns
for each country on world excess stock market returns, the estimated
$$R^2$$ values ranged from 14% to 52% across the 10 countries.
The highest $$R^2$$ values were found in Switzerland and the
Netherlands that are more international in scope than the other national
markets. National beta estimates ranged from 0.43 to 1.30 with an average
of 1.08. Hence, depending on the country, world equity market returns may
or may not explain national stock market returns.
Solnik ran a cross-sectional regression on an in-sample basis as follows:
$$\begin{aligned} \bar{R}_{ct}-\bar{R}_{ft}=\lambda _0 +
(6.6)
\lambda _w \hat{\beta }_i +u_{it}, \end{aligned}$$
where $$\bar{R}_{ct}-\bar{R}_{ft}$$ is the average excess return in
a country over the sample period, $$\hat{\beta }_c$$ is the estimated
beta in Eq. (6.5) for each country in the sample period (i.e.,
$$R_{it}$$ for an individual asset is changed to $$R_{ct}$$ for
the average stock market return for each country), and
$$\lambda _w$$ is the estimated risk premium for international beta
risk. He obtained $$\lambda _w=0.26$$ with a highly significant t-
statistic equal to 6.2. Since this is a two-week market risk premium, on an
annual basis, it is fairly large at approximately 6.5% per year. These results
strongly supported the IAPM. Further cross-sectional tests of individual
stocks in the ten countries were conducted, but the results did not suggest
that international market risk is significantly priced. To explain this finding,
he cited Black et al. (1972), who recommended that portfolios should be
used in cross-sectional tests to reduce or mitigate errors-in-the-variables
problems in individual stock beta estimates. From these findings Solnik
(1974b, p. 552) concluded that:

Although stock prices of each country tend to move together, the

international capital market seems to be sufficiently integrated and
efficient to induce an inter-national pricing of risk of common
stocks. While these results are in favor of the International Asset
Pricing Model, they indicate that more efforts should be devoted to
the study of international market structure in order to determine the
stochastic process followed by security prices around the world.

Heeding this recommendation, in tests of $$\alpha _0$$ in time-series

regression equation (6.5) using a global market factor and 13 industrialized
countries, Fama and French (1998) found that the single world factor IAPM
failed. The intercept term was significantly greater than zero, which
suggests that the model is incomplete in the sense that there is an error of
omission in terms of missing factors. Confirming this problem, when they
added a value factor (to be discussed in forthcoming Chapter 8) to the right-
hand-side of the IAPM, the intercept was reduced to near zero in most
cases. Consequently, they concluded that an “ ... international two-factor
model simply provides a parsimonious way to summarize the general
patterns in international returns.” (Fama and French 1998, p. 117) One
interpretation of this finding is that the value factor captures a state
variable per Merton’s ICAPM that is missing from the single factor IAPM
—that is, assuming that the world investment opportunity set is changing
over time in response to state variables, an international IAPM exists that is
comprised of a world market factor plus state variables or hedge
portfolio factors.
Extending the Fama and French study, Griffin (2002) tested domestic
versus global versions of the market model in the United States, United
Kingdom, Canada, and Japan. Like them, they supplemented the market
factor with other factors—namely, size and value factors based on firm-
level characteristics. Global factors were computed as value-weighted
averages of the domestic factors in the sample countries. Using stock
returns from 1981 to 1995, a major finding was that domestic factors
provided better goodness-of-fit to explain stock returns than global
factors in the United States, Canada, the United Kingdom, and Japan.
Because tests of the $$\alpha _0$$ intercepts in both domestic and
global models were significant in all but Japan, contrary to Fama and
French, they rejected both domestic and global model specifications that
contained both market and other factors.
Summarizing their findings, they inferred that stock markets around the
world are not integrated with respect to risk factors that explain expected
stock returns in different countries. In effect, contrary to Solnik’s IAPM,
stocks appear to be domestically rather than globally priced.4 Following this
study and other corroborating studies, researchers nowadays focus their
investigations of asset returns and pricing factors on the national rather that
global level. Apparently we will have to wait for world financial markets to
become more fully integrated before Solnik’s IAPM becomes a reality.

6.2.2 Exchange Rate Risk

Even though it is now generally accepted that assets are domestically (not
globally) priced in financial markets, it is possible that the changing values
of currencies can affect investors’ security returns and therefore their risks.
Imagine that you are a U.S. investor holding a European stock. In this case,
you hold two assets: (1) the European stock, and (2) euro currency. Your
return will be based on both the European stock return and the change in
currency values of the dollar vis-à-vis euro. To see this, consider the
following example.
1.
As a U.S. investor, you exchange 100 dollars for 80 euros at the
exchange rate 1 USD = 0.80 EUR or 1 EUR = 1.25 USD.
2.
You buy 10 shares of a European stock for 80 euros.
3.
One year later you sell the European stock for 100 euros. You earn a
stock rate of return equal to $$R_{iT} = (100 - 80)/80 = 25$$% over
period T.
4. Next, you exchange 100 euros for 130 dollars at the then exchange rate
 1 USD = 0.77 EUR or 1 EUR = 1.30 USD. You earn a euro rate of
return equal to $$R_{xT}$$ = (1.30 USD/EUR − 1.25
USD/EUR)/1.25 USD/EUR = 4.0%.
5.
Your total return is (1+RiT)(1+RxT) − 1 = 1.25 × 1.04 = 0.30, or 30 %,
which approximately equals the sum of the European stock return and
the euro return, or 25% + 4.0% = 29.0%.
This example shows that exchange rate movements affect international
investors’ rates of returns on foreign assets.
A related but different question is: Do exchange rate movements affect
the risk of the security investments themselves? That is, let’s say that you
are a U.S. investor and only buy domestic assets in the United States.
Should you be concerned about how exchange risk in currency markets
might affect the risks of your U.S. investments? Given that exchange rates
(e.g., USD/EUR) are more than 10 times more volatile than inflation
rates based on consumer and producer prices, and most investors are
concerned with inflation risk associated with potentially decreasing
purchasing power, it seems likely that exchange rate risk would be an
important risk factor.
To help answer the above question, Adler and Dumas (1984) defined
currency risk as the random movement of currency values over time. Over
100 currencies exist in the world, many of which trade in currency markets
on a daily basis. Indeed, the volume of currency trading on any given day
exceeds that of any other asset class, including stocks, bonds, commodities,
real estate, etc. The currency market is a large, inter-connected world
market that is primarily managed by global corporate banks. Currencies can
be classified as hard versus soft. Hard currencies are those of major
industrialized countries, including the dollar, euro, pound, yen, etc. Soft
currencies are those of developing countries and often are pegged to the
values of hard currencies. Currency risk, as measured by the volatility of
exchange rates of currencies over time, is generally higher for soft
compared to hard currencies. Even so, the values of hard currencies in
terms of exchange rates with other hard currencies can fluctuate
considerably over time. Hence, currency risk is substantial in world
currency markets.
 According to Adler and Dumas, exchange rate risk is a different but
related concept to currency risk that focuses on the sensitivity of a security
to currency movements. More specifically, how does currency risk affect
the risk of securities? Figure 6.3 illustrates the difference between currency
risk and exchange rate risk. A company that is exporting its goods benefits
from home currency depreciation as more foreigners would want to buy its
goods, and vice versa in the case of appreciation. Also, given home
currency appreciation, a goods importer experiences benefits from buying
at lower costs than previously, and vice versa with depreciation. Hence,
changes in exchange rates can impact the cash flows of exporting and
importing firms. Since stock returns should reflect all information (in an
efficient market) affecting the firm including the effects of exchange rate
movements, they proposed that exchange risk exposure can be estimated for
an investor in a home country by means of the following simple regression:

Fig. 6.3 Currency risk in world currency markets affects the stock returns in different countries
through their exchange rate risk exposure or sensitivity to national currency movements

$$\begin{aligned} R_{it}=\alpha _0 + \beta _{i,x} R_{xt}+\epsilon

(6.7)
_{it}, \end{aligned}$$
where $$R_{it}$$ is the return for the ith asset in the home country in
time period t (e.g., the dollar-denominated return for a security in the
United States), and $$R_{xt}$$ is the rate of return on the spot
exchange rate x in period t of dollars per euro, or [
$$USD / EUR _t - USD / EUR _{t-1}]/ USD / EUR _{t-1}$$.5
Notice that the exchange rate x is the home currency price of foreign
currency (or $$USD / EUR$$ here). Importantly, the
$$\beta _{i,x}$$ coefficient represents the total foreign exchange
exposure of the stock as measured by the elasticity of changes in stock
prices in response to a change in exchange rates. Adler and Dumas (1984,
p. 48) concluded:

What the regression coefficient concept of exposure can provide is a

single comprehensive measure that summarizes the sensitivity of the
 whole firm, as of a given future date, to all the various ways in
which exchange rate changes can affect it.

Jorion (1990) implemented empirical tests of exchange rate risk. To do this,

he modified the total exchange rate exposure equation (6.7) to take into
account the market factor as follows:
$$\begin{aligned} R_{it}=\alpha _0 + \beta _{i,m} R_{mt}+\beta
(6.8)
_{i,x} R_{xt}+\epsilon _{it}, \end{aligned}$$
where the proxy return for the market portfolio $$R_{mt}$$ is
augmented with exchange rate of return $$R_{xt}$$. He computed
$$R_{xt}$$ as the rate of return on the USD per unit of a basket of
currencies of 15 different countries (known as the Multilateral Exchange
Rate Model or MERM of the International Monetary Fund).6 Using U.S.
returns in the sample period 1971 to 1987 when exchange rates were
allowed to freely float in world currency markets, he found that most
exchange risk coefficients, or $$\beta _{i,x}$$, were insignificant. In
this period only 15-out-of-287 sample firms had significant exposure
coefficients at the five percent level. Looking more closely at 40 large
multinational firms that actively engage in international trade activities, he
detected significant exposure coefficients. Further tests confirmed that firms
with more foreign sales as a percentage of total sales had larger foreign
exchange risk exposure coefficients.
We can summarize the early empirical evidence on foreign exchange
risk and valuation as mixed. Most stocks’ returns do not appear to be
exposed to foreign exchange risk. However, the values of multinational
firms and firms with relatively high foreign sales are significantly affected
by foreign exchange risk. Many other researchers have investigated the
exposure of stocks to currency movements. Due to the lack of exposure for
most companies despite volatile movements in exchange rates over time,
many researchers consider this weak evidence to be one of the most
perplexing puzzles in modern asset pricing.

6.3 Consumption CAPM

Seminal work by Lucas (1978) and Breeden (1979) proposed a
consumption CAPM (CCAPM) version of the CAPM. Both authors
employed the ICAPM framework of Merton (1973). Instead of a market
beta related to the expected excess return on the market portfolio, a
consumption beta was proposed. Also, while the CAPM is a one-period
model based on the mean and variance of returns, the CCAPM is an
intertemporal model in which investors maximize expected lifetime utility.
Consumers buy and sell financial assets over time to smooth their
consumption of goods. For example, they sell (buy) assets when times are
bad (good) to maintain their consumption needs.
Lucas derived the CCAPM under simple assumptions about the
economy. There is one consumer (a representative agent) that seeks to
maximize their expected lifetime utility from one good. The consumer can
buy dividend-paying shares in companies that make the good in the
financial market to form a portfolio of financial assets. Stock prices
fluctuate over time with the state of the economy. The consumer seeks to
optimize their holdings of both the good and portfolio over time to
maximize their lifetime expected utility. A key insight is that the marginal
rate of substitution of current for future consumption by the consumer is
equivalent to the rate of return on a stock. This means that consumption is
equivalent to dividends over time.

Fig. 6.4 Consumers buy and sell financial assets over time to smooth consumption in the
consumption CAPM (CCAPM)

Lucas posited that intertemporal patterns of consumption and financial

asset transactions exist. The consumer will seek assets with negative
covariances with their consumption, which results in low risk premiums for
these assets. They will sell financial assets when the economy is in a
recession to supplement their consumption which has decreased due to
lower income at such times. Conversely, assets with positive covariances
with their consumption pattern (i.e., low values when consumption is lower)
will not be desired by investors and, therefore, will have higher risk
premiums to get people to hold them. In good economic times, consumers
will buy financial assets as incomes are higher to meet consumption needs.
Hence, as shown in Fig. 6.4, the lifetime pattern of consumption is
smoothed by means of buying and selling financial assets over time.
 In his model, the expected return on asset i at time t is generally
expressed as:
$$\begin{aligned} E({R}_{it})-R_{ft}=-\frac{1}{E[U({C}_{t})]}
(6.9)
Cov [U({C}_t), R_{it}], \end{aligned}$$
where $$E[U(C_t)]$$ is the expected utility from consumption (i.e.,
its expected value to the consumer), and $$Cov [U(C_t), R_{it}]$$ is
the covariance of consumption utility and asset returns related to systematic
risk that cannot be diversified away by the consumer. In this context,
$$\beta _i$$ is represented by the right-hand-side term. Assets with
negative covariance have higher risk premiums, and positive covariance
assets have lower risk premiums. Negative covariance assets offer insurance
against bad times and therefore tend to smooth consumption, whereas
positive covariance assets do not smooth consumption at such times.
Breeden (1979, p. 266) proposed a CCAPM that closely follows the
Lucas model:

... it is shown that Merton’s multi-beta pricing equation can be

collapsed into a single-beta equation, where the instantaneous
expected excess return on any security is proportional to its ‘beta’
(or covariance) with respect to aggregate consumption alone.

Since the CCAPM focuses on an easily identified variable (viz.,

aggregate consumption) rather than a number of unspecified variables (e.g.,
the difficult to proxy market portfolio in the CAPM), he argued that it can
be readily implemented and tested. Assuming there is an asset with returns
perfectly correlated with aggregate consumption over time, he specified
excess asset returns as follows:
$$\begin{aligned} E({R}_{it})-R_t=-\beta _C [E({R}_{Ct})-R_t],
(6.10)
\end{aligned}$$
where $$E(R_C)-R_t$$ is the excess return on aggregate
consumption of the perfectly correlated asset (e.g., over the riskless rate or
zero-beta portfolio rate), and $$\beta _C$$ is the consumption beta at
each instant in time t. As in Lucas, asset values are negatively related to
their covariance with respect to aggregate consumption. Higher (lower)
expected returns on assets are related to larger (smaller) covariances with
aggregate consumption. He contended that wealth in the CAPM was not a
sufficient measure of the marginal utility of a dollar to the investor. There
can be states of the world when wealth is low but the marginal utility of a
dollar could be low also. However, it is always true that, when the marginal
utility of a dollar is high, consumption is low in that state, and vice versa.
Due to this rationale, he claimed that the CCAPM using consumption beta
was more appropriate than the CAPM market beta in the determination of
equilibrium expected rates of return. Hence, from a theoretical perspective,
consumption beta is more justified than market beta.
According to Breedon, one potential problem in empirically testing the
CCAPM is that consumption is measured with error in the real world.
Despite this issue, he asserted that measured consumption reflects a greater
fraction of total consumption than proxies used for the market portfolio that
typically only include stocks (for example) to the exclusion of other asset
classes (e.g., bonds, commodities, real estate, etc.). In other words,
empirical proxies for aggregate consumption in the CCAPM enable better
tests of the theory than those used to test the CAPM. As discussed in
Chapters 4 and 5, consistent with the Roll (1977) critique, it could be that
evidence against the CAPM is due to bad proxies for the market portfolio.
Breeden et al. (1989) provided an early test of the CCAPM. In the
period 1959 to 1982, they measured aggregate consumption in two ways:
(1) quarterly changes in nondurable goods and services, and (2) a stock
portfolio with maximum correlation with the growth in aggregate
consumption. Also, the CAPM was comparatively tested using the
CRSP value-weighted index to proxy for the market portfolio. Their results
showed that market betas and consumption betas are highly correlated at
0.96 and 0.94 for the two consumption proxies, respectively.
Using estimated betas, the authors conducted cross-sectional Fama and
MacBeth (1973) tests of the CCAPM. Test assets included 17 portfolios
comprised of 12 industry stock portfolios, four bond portfolios, and the
CRSP value-weighted index. In summary, they found the following results.
1.
Plots of expected returns and estimated consumption betas revealed
relatively flat lines consistent with early tests of the CAPM. Given the
high correlation between market and consumption betas, this finding is
not surprising.
2. The significance of the market price of consumption risk (as measured
b ti t d $$\l bd C$$ ffi i t ti t d b t )
 by estimated $$\lambda _C$$ coefficients on estimated betas) was
mixed for different sample periods. In some samples, consumption risk
is priced, but in others it is not.
3.
Cross-sectional regression models had small estimated intercept terms
$$\lambda _0$$. This finding was interpreted as more consistent with
a riskless rate than a zero-beta portfolio borrowing rate.
4.
Both the CRSP value-weighted index and maximum correlation
portfolio with consumption (which were highly correlated with one
another) were not mean-variance efficient. This inference was based on
joint tests that rejected all $$\alpha _0$$ terms in a time-series
regression model equaled zero. Their results indicated weak cross-
sectional evidence for both the CAPM and CCAPM.
In later tests of the CCAPM by other researchers, similar findings have
been reported. For this reason, as new asset pricing models evolved,
especially multifactor models covered in forthcoming Chapters 8 and 9,
most researchers augment the market factor with other proposed factors. As
we will see, multifactor models were a major advance over the CAPM in
terms of their ability to explain average stock returns.

6.4 Production CAPM

Cochrane (1991) developed a production CAPM (PCAPM) that is based on
producers and production functions instead of consumers and consumption
functions in the CCAPM. He observed that stock returns are generally
accepted to be forward looking in that they forecast future economic
activity, including business investment and GNP growth. More specifically,
the PCAPM links asset returns to the production function—high expected
returns are explained by high business investment growth. Firms put to
work labor and capital to produce consumption goods. Similar to the logic
of the CCAPM, investment returns on capital (used in private firm
production) equal the return on a portfolio of assets over time (e.g., stock
returns). Figure 6.5 gives an example of this proposition with forecasts of
stock returns and capital investment returns that are highly correlated over
time. Thus, both the PCAPM and CCAPM connect real economic activity
(and associated macroeconomic risks) and expected stock returns. We can
think of the CCAPM as a demand side theory (consumers) and the PCAPM
as a supply side theory (producers). In equilibrium, given the law of one
price for each asset, these two theories are flip sides of the same coin.

Fig. 6.5 Close relation between stock return and capital investment forecasts in the production
CAPM (PCAPM)

A model of the form of CCAPM relation (6.10) can be specified to

represent the PCAPM. Rather than using the growth rate of aggregate
consumption, Cochrane constructed an investment/capital ratio (
$$I/k)_t$$ at time t. Quarterly data series were computed from 1947
to 1987. Based on a variety of tests, among other findings, he found a
positive correlation between forecasts of capital/investment ratios and stock
returns. This time-series evidence provided some support for the PCAPM.
In a second paper, Cochrane (1996) published more in-depth cross-
sectional asset pricing tests of the PCAPM. Because the tests did not
explicitly model a pure production model, he referred to the empirical
specification as an investment-based model. In this indirect test of the
PCAPM, he hypothesized the following single factor model:
$$\begin{aligned} R_{it}-R_{ft}=\alpha _0+\beta _i (R_{It}-
(6.11)
R_{ft})+\epsilon _{it}, \end{aligned}$$
where $$R_{It}$$ is the investment return equal to the extra output
sold by the firm upon investing in an additional unit of capital,
$$R_{ft}$$ is the riskless rate, and $$\beta _i$$ is the
investment beta. Here investment returns represent an asset pricing factor
for asset returns. His analyses employed investment returns for both
nonresidential and residential gross investments on a quarterly basis in the
United States. Using 10 portfolios sorted into size deciles as test assets,
cross-sectional regression $$\lambda$$ estimates were not significant
for the nonresidential investments factor but negatively significant for the
residential investments factor.
Further tests conditioned these investment factors on market variables—
namely, the term premium and dividend/price ratio of a stock market index
—to allow for time variation in the beta loadings or coefficient estimates
over time. In the next subsection, we overview conditional asset pricing
models. On a conditional basis, both investment factors were significant.
Also, he tested the CAPM and CCAPM and found that the investment-
based CAPM performed similar to the CAPM but outperformed the
CCAPM.7 Cochrane (1996, p. 615) concluded that:

The simple investment return model performs surprisingly well. The

investment return factors significantly price assets, the model is not
rejected, and it is able to explain a wide spread in expected returns
...

The PCAPM draws an important link between investment in the real sector
of the economy and asset returns in the financial sector. The fact that the
conditional model improved its performance suggests that risk parameters
are time-varying over the business cycle. Good and bad economic times are
relevant to asset pricing. We next take a closer look at conditional
models and their empirical setup.

6.5 Conditional CAPM

The CAPM and its companion models are unconditional models that
normally assume that expected returns are constant, the market factor and
other factors can be measured, and assets’ betas are fairly constant (or
stationary) over time. Cross-sectional tests of asset pricing models regress
unconditional expected returns on unconditional betas. However,
considerable research has shown that expected returns and estimated betas
of stocks vary considerably over time. That is, beta moves up and down in
the financial market. Why does this happen? One explanation is that beta
risk could change over the business cycle. In a recession, when economic
activity is depressed and the risk of business failure is higher than normal,
beta risks of companies prone to bankruptcy (e.g., due to high debt on their
balance sheet) are likely to be higher than otherwise.
The unconditional CAPM implies that expected excess returns are
proportional to market beta:
$$\begin{aligned} E({r}_{it}) = \frac{E({r}_{mt})}{\sigma
(6.12)
^2({r}_{mt})} Cov ({r}_{it}, {R}_{mt}), \end{aligned}$$
where lowercase r denotes excess returns $$E({R}_{it})-R_{ft}$$,
and the familiar unconditional beta is defined as
$$\beta _i= Cov ({R}_{it}, {R}_{mt})/\sigma ^2({R}_{mt})$$.
Recognizing the contributions of numerous authors, including Hansen and
Hodrick (1983), Gibbons and Ferson (1985), Ferson et al. (1987), Harvey
(1989), and Shanken (1990), we follow Harvey by writing the conditional
form of this expression as follows:
$$\begin{aligned} E({r}_{it} \vert I_{t-1})=\frac{E({r}_{mt}|I_{t-
1})}{\sigma ^2({r}_{mt}|I_{t-1})} Cov ({r}_{it}, {r}_{mt} \vert (6.13)
I_{t-1}), \end{aligned}$$
where expected excess returns and the covariance of excess returns with
excess market returns are conditional on previous information
$$I_{t-1}$$ in the financial markets, and the conditional beta is
$$\beta _{it} = Cov [{r}_{it}, {r}_{mt} \vert I_{t-1}]/\sigma
^2({r}_{mt}|I_{t-1})$$
. In line with this logic, a conditional CAPM can be specified as:
$$\begin{aligned} E({r}_{it} \vert I_{t-1}) = \beta _{it}
(6.14)
E({r}_{mt}|I_{t-1}), \end{aligned}$$
where $$\beta _{it}$$ has a time subscript to indicate variation over
time.
Following Cochrane (1996) discussed in the previous section,8 factors
can be scaled with instruments that are related to financial market variables.
This instrumental approach is consistent with the ICAPM of Merton, who
proposed that the investment opportunity set of investors changes over time
conditional on time-varying state variables. For example, we can scale the
market factor (or any risk factor) by specifying a conditional (time-varying)
beta as a linear function9:
$$\begin{aligned} \beta _{it} = \beta _{0i} + \beta _{1i} I_{t-1},
(6.15)
\end{aligned}$$
where $$\beta _0$$ is the unconditional CAPM beta, and
$$I_{t-1}$$ is a time-varying financial market variable that scales
beta to different values over time. The lagged variable $$I_{t-1}$$ is
known as an instrument to help capture the time-varying conditional beta. If
we also assume that the intercept term is conditional, so that
 $$\alpha _{it} = \alpha _{0i} + \alpha _{1i} I_t$$, the conditional
form of the market model can be estimated as:
$$\begin{aligned} r_{it}= (\alpha _{0i} + \alpha _{1i} I_{t-1}) +
(\beta _{0i} + \beta _{1i} I_{t-1}) r_{mt}+\epsilon _{it} (6.16)
\end{aligned}$$

$$\begin{aligned} r_{it}= \alpha _{0i} + \alpha _{1i} I_{t-1} +

\beta _{0i} r_{mt} + \beta _{1i} (r_{mt} I_{t-1})+\epsilon _{it}. (6.17)
\end{aligned}$$
Here $$\alpha _{0i}$$ and $$\beta _{0i}$$ are the
unconditional alpha and beta denoted earlier as $$\alpha _i$$ and
$$\beta _i$$, respectively. The new intercept
$$\alpha _1 I_{t-1}$$ varies over time with $$I_{t-1}$$. Also,
the new beta estimate $$\beta _{1i}$$ relates excess turns to the
interaction variable $$r_{mt} I_{t-1}$$, which is computed as excess
market returns $$r_{mt}$$ times variable $$I_{t-1}$$ at time t.
In effect, the conditional approach changes the single factor market
model into a multifactor model with a market factor plus an interaction
factor. As an example, Petkova and Zhang (2005) used a variety of financial
market variables, including default premiums on bonds, dividend yields,
Treasury bill rates, and the term premiums on bonds, as instruments in a
conditional CAPM. In this regard, any variable that could affect market risk
over time is a feasible instrument.
Does time variation in beta explain the weak evidence for the CAPM?
Research evidence appears to be mixed in this regard. For example, Lettau
and Ludvigson (2001) compared both unconditional and conditional
versions of the CAPM and CCAPM to the three-factor model of Fama and
French (1992, 1993, 1995). They instrumented betas with the aggregate
consumption/wealth ratio in the United States computed quarterly from
1968 to 1998. The three-factor model is a multifactor model covered in
Chapter 8 that has proven to better explain average stock returns than the
CAPM. It is comprised of market, size, and value factors, where the latter
two factors are based on firm characteristics. In cross-sectional tests, the
authors found that the conditional CCAPM performed as well as the three-
factor model, which supports the conditional approach.
 By contrast, other research by Lewellen and Nagel (2006) did not find
support for the conditional CAPM. Since it is not known what instruments a
researcher should use in the conditional CAPM, they estimated beta over a
large number of short intervals of time to capture its time variation and that
of alpha terms also. Using U.S. stock returns from 1964 to 2001, they
reported that betas change over time but not sufficiently to explain weak
evidence for the CAPM. Also, beta variation over time could not account
for the size and value factors of Fama and French or the momentum
factor is to be discussed in Chapter 8.

Fig. 6.6 Beta estimates for the market factor tend to vary randomly over time consistent with the
unconditional CAPM but contrary to the conditional CAPM

Figure 6.6 gives an illustration of beta for a factor varying considerably

over time in a random fashion. Here beta is not moving in a regular pattern
that would suggest an underlying state variable (e.g., macroeconomy, firm
characteristic, etc.). Consistent with this observation, Lewellen and
Nagel found that the conditional CAPM did not improve upon the
unconditional CAPM in time-series intercept ( $$\alpha _0$$) tests as
well as modifications of Lettau and Ludvigson cross-sectional tests. Similar
inferences were made in comparisons of the conditional and unconditional
consumption CAPM (CCAPM). Because conditional alphas (
$$\alpha _0$$s) were large and significant, they concluded that the
conditional CAPM did not hold.10 This evidence suggests that business
cycle effects in asset pricing models are not sufficiently large to invalidate
unconditional models.
Nowadays, most asset pricing studies do not incorporate the business
cycle in their analyses. Long periods of time are used that contain a number
business cycles with economic expansions and recessions. Of course,
results of different models can vary over time and this may indeed be due to
the influence of a business cycle in the sample period used by a researcher.
Nonetheless, regardless of the economic state, a valid asset pricing model
should do a good job of measuring returns with respect to various
systematic risks.
6.6 Summary
Early evidence that weakly supported the CAPM sparked researchers to
develop alternative specifications in an effort to be more compatible with
real-world conditions in capital markets. Merton’s (1973) famous
intertemporal CAPM (ICAPM) extended the single-period or static CAPM
to multiple periods. Investors make decisions about their future wealth over
their lifetime and even beyond to their heirs. In this more realistic situation,
the investment opportunity set facing investors can change over time. He
believed that state variables explain the shifting opportunity set, which itself
introduces new risks for investors. Investors can form hedge portfolios to
manage risks associated with state variables. These portfolios are factors
not taken into account by the single market factor CAPM. His three-fund
theorem introduced a third asset or portfolio that is perfectly negatively
correlated with the riskless rate (i.e., the CAPM’s two-fund theorem
includes the market portfolio and riskless asset). Practically speaking, the
ICAPM can be empirically modeled by adding multifactors beyond the
market factor to the market model. Unfortunately, the ICAPM does not
identify the state variables that investors use in the real world to price
assets.
Merton’s landmark ICAPM stimulated the development of CAPM
variants. The international asset pricing model (IAPM) of Solnik (1974a,
1974b) broadened the scope of the CAPM from a single national market to
an integrated world market with inter-connected national markets. Now
investors face exchange rate risk associated with currency movements that
was not present in the CAPM. Solnik’s IAPM specifies excess returns of
assets as a function of beta times of the world stock market premium over
the average interest rate on a global portfolio of bonds. Cross-sectional tests
of 10 countries supported the IAPM in terms of a significant world market
price of risk (or $$\lambda _w$$). He concluded that national stock
markets are integrated globally. However, later evidence by Fama and
French (1998) and Griffin (2002) did not support the IAPM due to
significant $$\alpha _0$$ intercept terms greater than zero in the time-
series regression estimation of the empirical IAPM. Even though national
stock markets are not internationally integrated, and asset pricing is
therefore domestic (or local) rather than global in nature, currency risk
related to volatile exchange rates can affect stock returns. Adler and
Dumas (1984) advocated that exchange rate exposure of stocks can
influence returns. Empirical tests by Jorion (1990) confirmed this
conjecture but with mixed evidence. Primarily stocks of large multinational
firms engaged in international trade activities were affected by exchange
rate movements but not most other stocks. Due to the general insensitivity
of most stocks to fairly volatile currency movements over time, an
exchange rate risk puzzle persists.
Another extension by Lucas (1978) and Breeden (1979) used the
ICAPM to derive a consumption CAPM (CCAPM). In this model, beta is
related to consumption rather than the market portfolio in the CAPM.
Investors seek to maximize their expected lifetime utility of consumption.
By buying and selling financial assets, they smooth their consumption over
time. This consumption smoothing behavior causes investors to prefer
assets with negative covariances between returns and consumption (which
benefits them in bad states such as an economic recession), thereby
reducing risk premiums on these assets. By contrast, positive covariance
assets have relatively higher risk premiums. In general, systematic risk for
assets is dependent on the state of the economy over time and related
consumption behavior. Despite some empirical advantages of the CCAPM,
such as the ability to obtain better proxies for aggregate consumption than
the CAPM’s market portfolio, stock market tests by Breeden et al.
(1989) showed that the CCAPM did not provide a better fit to returns than
the CAPM. Again, a relatively flat relation between expected returns and
estimated consumption betas was obtained. This finding is not surprising in
view of the very high correlation between market betas and consumption
betas.
Related to the CCAPM with consumers and consumption functions,
Cochrane (1991) proposed the production CAPM (PCAPM) based on
producers and production functions. Stock returns are forward-looking
forecasts of future economic activity and the related production function.
While the CCAPM focuses on the supply side (consumers) in the economy,
the PCAPM takes a demand side (producers) perspective. He showed that
the investment/capital ratio is positively correlated with stock returns. Also,
he tested an investment-based version of the PCAPM that performed
similar to (better than) the CAPM (CCAPM) in cross-sectional tests.
Lastly, the conditional CAPM was developed by numerous authors to
take into account time-varying risk parameters in asset pricing models (or
betas). In this regard, it is reasonable to believe that risk parameters will
change in response to the business cycle with economic expansions and
recessions. To model time-varying risk parameters, lagged instrumental
variables are introduced that model beta coefficients as a function of
financial market variables. In this way, the unconditional CAPM is
extended to conditional state variables consistent with the ICAPM.
Evidence by Lettau and Ludvigson (2001) supported the conditional
CAPM in cross-sectional tests. However, Lewellen and Nagel (2006) later
conducted direct tests of betas in short intervals of time and found that betas
did not change sufficiently over time to justify the conditional CAPM.
Overall, the ICAPM and its derivative CAPMs have enhanced our
understanding of asset pricing. No doubt these models will play an
important role in future research and development.

Questions
1.
Merton (1973) developed the intertemporal CAPM (ICAPM). What is
different about the ICAPM compared to the CAPM? What new
assumption does it make?
2.
In the ICAPM, how do investors respond to unfavorable shifts in
future investment opportunities? What are state variables?
3.
In the ICAPM, what is the three-fund theorem? In view of this
theorem, write the equation for the ICAPM.
4.
In the ICAPM, if correlations $$\rho _{iN} = 0$$ for all securities i,
which also implies that $$\rho _{NM} = 0$$ (where M is the
market portfolio), but $$r_N$$ is perfectly negatively correlated
with $$R_f$$ (which is true by definition), the correlations of
stocks with the future riskless rate would be zero. What happens to the
ICAPM is this special case? What is the beta of security N in this
case? Is N related to the zero-beta portfolio of Black in this case? If
the market beta $$\beta _{i,M} = 0$$ but $$\beta _{i,N}&gt;0$$,
is the expected return on the i asset equal to the riskless rate?
5. Solnik (1974) proposed an international asset pricing model (IAPM)
based on an international three-fund theorem. What is this theorem
 based on an international three fund theorem. What is this theorem
and how does it change the CAPM? Write the equation for the IAPM
and discuss Solnik’s model.
6.
How did Adler and Dumas (1984) define currency risk? How is
currency risk different than exchange rate risk? How did they propose
to measure the exchange risk exposure of a security? Write an
equation that can be used to measure this exposure.

7.
Jorion (1990) investigated the exchange risk exposure of U.S. stocks.
He modified the total exchange risk exposure equation of Adler and
Dumas (1984) in what way? What did he find?
8.
Lucas (1978) and Breeden (1979) proposed the consumption CAPM
(CCAPM) using the ICAPM framework of Merton (1973). In this
model, consumers buy and sell financial assets over time to smooth
consumption and maximize lifetime expected utility. What assets will
consumers seek? Why? How does it affect the risk premiums on these
assets? What happens to other assets?
9.
Breeden (1979) specified a CCAPM that collapses the multi-beta
Merton (1973) ICAPM into a single-beta equation. Write this model
and define its variables and parameters. Why did he believe that the
CCAPM was a better model than the CAPM from a theoretical
perspective as well as in terms of empirical testing? Later work by
Breeden et al. (1989) conducted empirical tests of the CCAPM. What
did they find?
10.
Cochrane (1991, 1996) created a production CAPM (PCAPM). How
is the PCAPM different from the consumption CAPM (CCAPM)? He
tested it using a single factor model. Write the equation for this model
and discuss his results. What did he conclude?
11. Assume that we scale the market factor with an instrumental lagged
variable denoted as $$I_{t-1}$$. Write a conditional (time-varying)
linear function for beta and define its terms. Assuming that the alpha
term is conditional also, incorporate this instrument into the market
model and discuss the parameters in this conditional form of the
market model.
12.
Does time-variation in beta explain the weak evidence for the CAPM?
Discuss empirical results of studies testing conditional CAPMs.

Problems
1.
Write the ICAPM as a multifactor model with state variables. What
could be used to proxy for the state variables?
2.
Write the market model form of Solnik’s IAPM. Solnik ran this model
on 10 major industrialized countries from 1966 to 1971. What did he
find? What did cross-sectional regression tests find? What did later
authors find in testing the IAPM?
3.
Assume that you are a U.S. investor buying a European stock:
1.
You exchange 100 dollars for 90 euros at the exchange rate 1 USD
= 0.90 EUR or 1 EUR = 1.11 USD.
2.
You buy 10 shares of a European stock for 90 euros.
3.
One year later you sell the European stock for 100 euros. You
exchange the 100 euros for 120 dollars at the exchange rate 1 USD
= 0.83 EUR or 1 EUR = 1.20 USD.
What is your stock rate of return? What is your euro rate of return?
What is your total return?

References
Adler, M., and D. Bernard. 1984. Exposure to currency risk: Definition and measurement. Financial
Management 13: 41–50.

Black, F. 1972. Capital market equilibrium with restricted borrowing. Journal of Business 45: 444–
454.
[Crossref]

Black, F., M.C. Jensen, and M. Scholes. 1972. The capital asset pricing model: Some empirical tests.
In Studies in the Theory of Capital Markets, ed. M.C. Jensen. New York, NY: Praeger.
Breeden, N.D. 1979. An intertemporal asset pricing model with stochastic consumption and
investment opportunities. Journal of Financial Economics 7: 265–296.
[Crossref]

Breeden, N.D., M.R. Gibbons, and R.H. Litzenberger. 1989. Empirical tests of the consumption-
oriented CAPM. Journal of Finance 44: 231–262.

Chen, N., R. Roll, and S.A. Ross. 1986. Economic forces and the stock market. Journal of Business
59: 383–404.
[Crossref]

Cochrane, J.H. 1991. Production-based asset pricing and the link between stock returns and
economic fluctuations. Journal of Finance 46: 209–237.
[Crossref]

Cochrane, J.H. 1996. A cross-sectional test of an investment-based asset pricing model. Journal of
Political Economy 104: 572–621.
[Crossref]

Cochrane, J.H. 2005. Asset Pricing. Revised. Princeton, NJ: Princeton University Press.

Fama, E. F. 1970. Multiperiod consumption-investment decisions. American Economic Review 60:

163–174.

Fama, E.F., and K.R. French. 1992. The cross-section of expected stock returns. Journal of Finance
47: 427–465.
[Crossref]

Fama, E.F., and K.R. French. 1993. The cross-section of expected returns. Journal of Financial
Economics 33: 3–56.
[Crossref]

Fama, E.F., and K.R. French. 1995. Size and book-to-market factors in earnings and returns. Journal
of Finance 50: 131–156.
[Crossref]

Fama, E.F., and K.R. French. 1998. Value versus growth: The international evidence. Journal of
Finance 53: 1975–1999.
[Crossref]

Fama, E.F., and J.D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy 81: 607–636.
[Crossref]

Jorion, P. 1990. The exchange-rate exposure of U.S. multinationals. Journal of Business 63: 331–
345.
[Crossref]

Ferson, W.E., S. Kandel, and R.F. Stambaugh. 1987. Tests of asset pricing with time-varying
expected risk premiums and market betas. Journal of Finance 42: 201–220.
[Crossref]
Gibbons, M.R., and W.E. Ferson. 1985. Testing asset pricing models with changing expectations and
an unobservable market portfolio. Journal of Financial Economics 14: 217–236.
[Crossref]

Griffin, J.M. 2002. Are the Fama French factors global or country specific? Review of Financial
Studies 15: 783–803.
[Crossref]

Hansen, L.P., and R.J. Hodrick. 1983. Risk averse speculation in the forward foreign exchange
market: An econometric analysis of linear models. In Exchange Rates and Iinternational
Macroeconomics, ed. J.A. Frenkel. Chicago, IL: University of Chicago Press.

Harvey, C.R. 1989. Time-varying conditional covariances in tests of asset pricing models. Journal of
Financial Economics 24: 289–317.
[Crossref]

Karolyi, G.A., and R.M. Stulz. 2002. Are financial assets priced locally or globally?’’. In The
Handbook of the Economics of Finance, ed. G. Constantinides, M. Harris, and R. Stulz. Amsterdam,
Netherlands: North Holland Publishing Co.

Kolari, J.W., W. Liu, and J.Z. Huang. 2021. A New Model of Capital Asset Prices: Theory and
Evidence. Cham, Switzerland: Palgrave Macmillan.
[Crossref]

Lettau, M., and S. Ludvigson. 2001. Resurrecting the (C)CAPM: A cross-sectional test when risk
premia are time-varying. Journal of Political Economy 109: 1238–1287.
[Crossref]

Lewellen, J., and S. Nagel. 2006. The conditional CAPM does not explain asset pricing anomalies.
Journal of Financial Economics 82: 289–413.
[Crossref]

Lucas, R.E. 1978. Asset prices in an exchange economy. Econometrica 46: 1429–1445.
[Crossref]

Merton, R.C. 1973. An intertemporal capital asset pricing model. Econometrica 41: 867–887.
[Crossref]

Petkova, R., and L. Zhang. 2005. Is value riskier than growth? Journal of Financial Economics 78:
187–202.
[Crossref]

Roll, R. 1977. A critique of the asset pricing theory’s tests, part I: On past and potential future
testability of the theory. Journal of Financial Economics 4: 129–176.
[Crossref]

Ross, S.A. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13: 341–
360.
[Crossref]

Shanken, J. 1990. Intertemporal asset pricing: An empirical investigation. Journal of Econometrics

45: 99–120.
[Crossref]

Simin, T. 2008. The poor performance of asset pricing models. Journal of Financial and Quantitative
Analysis 43: 335–380.
[Crossref]

Solnik, B.H. 1974a. The international pricing of risk: An empirical investigation of the world capital
market structure. Journal of Finance 29: 365–378.
[Crossref]

Solnik, B.H. 1974b. An international market model of security price behavior. Journal of Financial
and Quantitative Analysis 9: 537–554.
[Crossref]

Footnotes
1 Merton cited Fama (1970), who had observed that intertemporal portfolio maximization can be
treated as if the investor had a single-period utility function.

2 Browning motion in physics is the random motion of particles.

3 See equation (34) in Merton (1973, p. 882). The derivation of this model is quite complicated and
beyond the scope of the present text. He utilized stochastic differential calculus, complex continuous
functions, instantaneous returns (as opposed to discrete one-period returns), and optimality
conditions to solve the asset pricing problem.

4 For an excellent survey of the issue of domestic (local) versus global asset pricing, see Karolyi and
Stulz (2002).

5 If $$R_{xt}&gt;0$$, we can infer a depreciation of the dollar against the euro, as each euro can
buy more dollars at time t than before at time $$t-1$$. If $$R_{xt}&lt;0$$, then the dollar
appreciated against the euro.

6 A more widely used currency basket by the IMF is the Special Drawing Right (SDR) consisting of
currencies in the United States, Europe, Japan, and the United Kingdom. The Federal Reserve in the
United States constructs a number of other currency baskets. In general, currency baskets are formed
using trade weights of the different exchange rates.
7 Also, he tested a five-factor model by Chen et al. (1986) that contains the growth in industrial
production and some inflation and interest rate variables. Like the CAPM, the performance of this
multifactor model was similar to the PCAPM.

8 See also Cochrane (2005, Section 8.4).

9 Beyond the scope of the present discussion, a first-order Taylor series can be used to approximate
the potentially complex effects of an instrument on factor loadings. Taylor series are well known in
the sciences and represent an infinite sum of derivatives of a mathematical function at a single point.

10 Simin (2008) also found that conditional versions of the CAPM and Fama and French three-
factor model did not outperform their unconditional versions in terms of one-month-ahead cross-
sectional tests of their predictive ability.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_7

7. The Arbitrage Pricing Theory Model

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_7

One of the most controversial assumptions of the CAPM is that returns are
normally distributed, which implies that agents have quadratic utility
functions. This assumption guarantees the mean-variance efficiency of the
market portfolio. Also, it leads to a linear equilibrium relation between the
expected return on an asset and its sensitivity to the expected market
premium (i.e., beta risk). This implied linear relation is widely accepted due
to its sound intuition and easy empirical specification.
Taking linearity as a starting premise, Ross (1971, 1974, 1976)
developed the arbitrage pricing theory (APT).1 The APT depends on no-
arbitrage conditions in the financial market. The underlying intuition is that
the total variation of the return on a single asset stems from a (small)
number of common factors and a random idiosyncratic residual term.
Although the APT is based on the similar intuition of linearity in the
CAPM, it is much more general. This chapter discusses briefly the concept
of no arbitrage, introduces the basic idea of Ross’s APT model, and reviews
empirical tests of the model. Huberman and Wang (2017) give an excellent
overview of the APT and its practical applications.

7.1 APT Model

7.1.1 Assumptions
The key assumption of the arbitrage pricing theory (APT) is the absence of
arbitrage opportunities. Another important assumption is that there are
common risk factors that are linearly related to individual asset returns.
Other assumptions, including the normality of returns that implies investors
have quadratic utility functions, are similar to the CAPM. Summarizing
these starting conditions, the APT assumes : (1) risk-averse investors; (2)
investors can hold long and short positions including borrowing and lending
at the same rate; and (3) capital markets are perfectly competitive and
frictionless. The most important distinction between the APT and CAPM is
replacing the CAPM assumption of homogeneous expectations about the
distribution of assets’ returns with homogeneous beliefs about the linear
factor model of returns. Because the CAPM leads to a linear relationship
between expected returns and the market portfolio with completely
diversifiable return-specific idiosyncratic risk, the CAPM meets the no-
arbitrage condition of the APT and, therefore, can be considered as a
special case of APT.

7.1.2 Theoretical Model

We can introduce the basic idea of the APT model in terms of a single
factor model with the following return generating process:
$$\begin{aligned} R_i = E(R_i) + \beta _i F + \epsilon _i,
(7.1)
\end{aligned}$$
where $$R_i$$ is the random rate of return on the ith asset, $$E(R_i)$$
is the expected return on the ith asset, $$\beta _i$$ is the sensitivity of the
ith asset’s return with respect to a factor, F is a random common factor with
respect to all returns with $$E(F) = 0$$, and $$\epsilon _i$$ is the
random error term idiosyncratic to the ith asset that is not correlated with F
or the error terms of other assets, and $$i = 1, \ldots , n$$ assets which is
assumed to be a large number.

Fig. 7.1 Arbitraging riskless profits by buying and selling mispriced assets
(Source Adapted from Wikipedia [https://en.wikipedia.org/wiki/Arbitrage_pricing_theory])

The no-arbitrage condition has two main features: (1) no wealth; and (2)
no risk must earn zero return on average. These features are easy to
understand in terms of riskless returns. If $$R_f$$ and
$$R_f^*$$ are two riskless rates that differ, then borrowing at the
lower rate and investing in the higher rate would generate a possibility to
earn unlimited riskless returns generated by the interest rate difference. As
another example, Fig. 7.1 shows how to arbitrage riskless profits by buying
underpriced assets and selling overpriced assets.
Arbitrage portfolios are portfolios that involve no wealth, which is
known as zero-investment portfolios. In this respect, buying assets in a long
portfolio is financed by selling short other assets (i.e., selling borrowed
assets). Mathematically, such a zero-investment portfolio satisfies the
following condition:
$$\begin{aligned} \sum _{i = 1}^n w_i = 0, \end{aligned}$$ (7.2)
where $$w_i$$ are the amounts invested in asset i.
In developing the APT it is assumed that the portfolio weights
$$w_i$$ are of order 1/n to guarantee a well-diversified zero-
investment portfolio. This assumption allows us to utilize the law of large
numbers in probability theory to eliminate idiosyncratic risks from the
portfolio. Given that returns to satisfy the representation in Eq. (7.1), the
arbitrage portfolio becomes
$$\begin{aligned} R_p= &amp; {} \sum _{i = 1}^n w_iR_i \\
\nonumber= &amp; {} \sum _{i = 1}^n w_i E(R_i) + \sum _{i = 1}^n (7.3)
w_i \beta _i F + \sum _{i = 1}^n w_i \epsilon _i. \end{aligned}$$
In order to make this a zero risk arbitrage portfolio, we should be able to
eliminate the last two terms in the second line of the equation. That is, the
systematic risk component $$\sum _i w_i\beta _i F$$ and the residual
idiosyncratic risk $$\sum _iw_i\epsilon _i$$ drop out. The latter is
eliminated by the independence assumptions of random error terms
$$\epsilon _i$$ from common factor F and the error terms of different
assets. These assumptions combined with the assumption that
$$w_i \approx \pm 1/n$$ imply that the last term is zero, or
$$\sum _i w_i \epsilon _i \approx 0$$ for large n as idiosyncratic
risks are diversified away. Subsequently, for large n we can discard the error
from (7.3) and approximate the portfolio by
$$\begin{aligned} R_p = \sum _{i = 1}^n w_iE(R_i) + \sum _{i =
(7.4)
1}^n w_i \beta _i F. \end{aligned}$$
Note that $$\sum _i w_i\beta _i F = F \sum _i w_i \beta _i$$. Under
the zero-investment restriction in Eq. (7.2), we can always select weights
$$w_i$$ such that
$$\begin{aligned} \sum _i w_i\beta _i = 0. \end{aligned}$$ (7.5)
With these weights, all the risks from the arbitrage portfolio become
eliminated so that
$$\begin{aligned} R_p = \sum _{i = 1}^n w_i E(R_i),
(7.6)
\end{aligned}$$
which is a riskless, zero-investment portfolio. Therefore, because the zero-
investment portfolio is riskless, the no-arbitrage condition implies that we
must have $$R_p = 0$$, or
$$\begin{aligned} \sum _{i = 1}^nw_i E(R_i) = 0. \end{aligned}$$ (7.7)
Given that this must hold for all zero-investment portfolios with
$$\sum _i w_i = 0$$ and $$\sum _i w_i\beta _i = 0$$, pure
linear algebra implies that the expected return, or $$E(R_i)$$, is a
linear combination of a constant and the asset’s beta. This insight yields the
APT model for the expected return on each asset of the form:
$$\begin{aligned} E(R_i) = \gamma _0 + \gamma _1 \beta _i,
(7.8)
\end{aligned}$$
where $$\gamma _0$$ and $$\gamma _1$$ are common to all
stocks, and $$\beta _i$$ is the stock’s sensitivity to the risk factor.
It should be noted that, because all arbitrage portfolios are not
necessarily well diversified, Eq. (7.8) may hold only approximately. Taking
the approximation as a starting point, Huberman (1982) provided an
interesting approach to derive the APT model. Using purely the no-
arbitrage argument and without explicit reference to the meaning of well
diversified, we can utilize linear algebra to write:
$$\begin{aligned} E(R_i) = \gamma _0 + \gamma _1 \beta _i +
(7.9)
\alpha _i \end{aligned}$$
wherein alpha ( $$\alpha _i$$) is orthogonal to both the vector of ones
and betas. Huberman and Wang (2017) refer to Eq. (7.9) as an approximate
APT model and Eq. (7.8) as an exact APT model. Even though the
representation in Eq. (7.9) is purely a mathematical identity, the way it is
constructed makes $$\alpha _i$$ satisfy the zero-investment arbitrage
portfolio restrictions in Eqs. (7.2) and (7.5), which can be utilized as a basis
for defining a zero-investment portfolio. With this convention, Huberman
derived the main result of the APT theory that, if there is no arbitrage, the
deviations, or
$$\alpha _i = E(R_i) - (\gamma _0 + \gamma _i \beta _i)$$, must be
negligible as the number of stocks n grows, i.e., the representation of the
expected return $$E(R_i)$$ in Eq. (7.8) is essentially accurate.2 Of
course, from a financial economics perspective, the $$\alpha _i$$
parameter in equation (7.9) can be interpreted as a pricing error.
These results must hold for all assets. In particular, if there exists a
riskless asset with return $$R_f$$ whose sensitivity to the common
risk factor F is zero by definition, Eq. (7.8) yields for the riskless asset
$$R_f = \gamma _0$$. Therefore, if a riskless rate exists, the APT
model becomes
$$\begin{aligned} E(R_i) - R_f = \gamma _1 \beta _i.
(7.10)
\end{aligned}$$
Alternatively, setting $$\lambda = \gamma _1 + R_f$$, we can re-
write Eq. (7.10) in the commonly expressed form
$$\begin{aligned} E(R_i) - R_f = (\lambda - R_f)\beta _i,
(7.11)
\end{aligned}$$
which is known as the APT model for asset returns. The difference
$$\lambda - R_f$$ is common to all stocks as the factor risk
premium, and $$\lambda$$ is the expected value of the common risk
factor. Finally, if F is the market portfolio, the APT model reduces to the
CAPM.
Example: Assume that the common factor in the APT model is the
market portfolio with expected return $$\lambda = E(R_m)$$. Let
$$R_p$$ be the return of a well-diversified portfolio such that
$$R_p^e = \alpha _p + \beta _p R_m^e,$$
where $$R_p^e = R_p - R_f$$ is the excess return of the portfolio,
and $$R_m^e = R_m - R_f$$ is the market excess return. If
$$\alpha _p \ne 0$$, there is an arbitrage opportunity. The general
recipe is that, if $$\alpha _p$$ is positive, short the market, and if
$$\alpha _p$$ is negative, short the portfolio.
More precisely, as
$$\alpha _p = R_p - \beta _p R_m - (1 - \beta _p)R_f$$, then if
$$\alpha _p &gt; 0$$, the sign in front of each return gives the long (
$$+$$) versus short (−) position. That is, for a desired riskless
arbitrage income of $$\alpha _p\times x$$ dollars, multiply both sides
by x. This would result in buying portfolio p with x dollars, selling short the
market portfolio with $$\beta _px$$ dollars, and borrowing/lending
$$(1 - \beta _p)\times x$$ dollars (borrowing if
$$\beta _p &lt; 1$$, lending if $$\beta _p &gt; 1$$, and neither
if $$\beta _p = 1$$). If $$\alpha _p &lt; 0$$, the strategy is
simply reversed.
The single factor model can be easily extended to multiple factors,
wherein the single factor F in Eq. (7.1) is replaced by K common
factors denoted $$F_1, F_2, \ldots , F_K$$ with slope coefficients
$$\beta _{i1}, \beta _{i2}, \ldots , \beta _{iK}$$, respectively. Each
factor is orthogonal to (or uncorrelated with) all other factors. Then the
multifactor model becomes:
$$\begin{aligned} R_i = E(R_i) + \beta _{i,1} F_1 + \cdots + \beta
(7.12)
_{i,K} F_K + \epsilon _i. \end{aligned}$$
In this case the equilibrium relation in Eq. (7.10) generalizes to the
following form:
(7.13)
 $$\begin{aligned} E(R_i) - R_f = \beta _{i,1}\gamma _1 + \cdots +
\beta _{i,K}\gamma _K, \end{aligned}$$
where $$\gamma _1, \ldots , \gamma _K$$ are risk premiums
associated with the K common factors.
The arbitrage portfolio is again selected to diversify away asset-specific
idiosyncratic risk. Similarly, the weights $$w_i$$ can be selected
similar to the single factor case such that
$$\sum _{i = 1}^nw_i\beta _{i,k} = 0$$ for all
$$k = 1, \ldots , K$$. Finally, it is notable that the APT applies even if
some of the error terms are correlated within some clusters of assets, like
industries, provided that the error terms are independent between the
clusters and the number of clusters is large. In this case the number of
clusters takes care of the diversification effect.
Unlike the CAPM mean-variance theory, which implies a linear
relation with the market portfolio, the APT neither identifies the number of
factors nor what the factors are. This generality is one of the main
weaknesses of the APT model. In the next section we discuss attempts by
researchers to empirically identify the number of factors in the APT.

7.2 Empirical Testing

Even though the market portfolio of the CAPM is unobservable, the
procedures for empirically testing the CAPM are fairly clear. Testing the
APT is more elusive as theory does not identify the number of factors or
their identity. That said, the Achilles’ heel of the CAPM is the
unobservability of the market portfolio. Because the APT does not require
the measurement of the market portfolio, Roll and Ross (1980) argued that
the APT is more readily testable. Even so, substantial debate surrounds the
testability of the APT. Shanken (1982) and Dybvig and Ross (1985) have
exchanged different views on the testability of the APT. Also, Dhrymes,
Friend, and Gultekin (1984) have questioned the empirical relevance of the
model.
In an effort to test the APT, Roll and Ross (1980) used daily U.S. stock
returns from July 3, 1962 to December 31, 1972. For computational reasons
to extract the underlying factors, return series were divided alphabetically
into 42 groups of 30 securities each. For each group: (1) a
 $$30\times 30$$ covariance matrix was computed from the return
time series; (2) a maximum likelihood factor analysis was used to estimate
the number of factors and the factor loadings; (3) and estimates of cross-
sectional regressions were used to measure the significance of risk
premiums associated with different estimated factors.
Before discussing their results, we should mention that factor analysis is
a statistical method that seeks to find the number of dimensions in sample
data observations. In our daily lives, we normally think in terms of length,
width, and depth dimensions. Of course, time is a fourth dimension. In the
present instance, the data are stock returns over time for 42 groups of
securities. We do not know how many risk dimensions are present within
these stock returns. The factor analysis seeks to identify how many factors
exist within the data. Each factor derived by the factor analysis corresponds
to an orthogonal risk dimension. Using statistical methods, the factor
analysis generates the factors in rank order from most to least important.
In their factor analysis, which was used for identifying and estimating
factors, the authors over-fitted the model to include a few extra factors.
Using these factors, they estimated a time-series regression model. Next,
based on the estimated loadings in the time-series model, they estimated
Fama and MacBeth (1973) cross-sectional regressions of the form:
$$\begin{aligned} \bar{R}_{it} - R_f = \gamma _{1t} \hat{\beta
}_{i1} + \cdots + \gamma _{Kt}\hat{\beta }_{iK} + u_{it}, (7.14)
\end{aligned}$$
where $$\hat{\beta }_{ij}$$ are estimated loadings corresponding to
the factors, $$i = 1, \ldots , n$$ (number of stocks),
$$j = 1, \ldots , K$$ (number of factors), $$\gamma _{jt}$$ is
the risk premium associated with factor j estimated at each time point
$$t = 1, \ldots , T$$, and $$u_{it}$$ is the regression error term.
The final estimates of the factor risk premiums, or $$\gamma _j$$,
$$j = 1, \ldots , K$$, were obtained by averaging the respective cross-
sectional estimates for the n stocks. Insignificance of the estimated
premiums of extra factors would serve to double check the inference of the
most likely number of the underlying true return generating factors
generated by the factor analysis.
Using these procedures, Roll and Ross determined that there were five
factors in the factor analysis. The two extra factors were found to be
insignificant in the cross-sectional regression tests. On the basis of their
empirical results, Roll and Ross concluded that: “... at least three factors are
important for pricing, but it is unlikely that more than four are present ...”
(Roll and Ross (1980, p. 1092). They concluded that the evidence supported
the APT.
While evidence of several priced factors is consistent with the APT,
Roll and Ross tested further whether there might be other variables, not
related to undiversifiable risk, that were priced. If such variables exist, the
APT would be rejected. To do this, the authors investigated the total
variance or “own” variance of individual stock returns. It is well known that
total variance is positively correlated with the sample mean return.3 The
initial results from cross-sectional regressions augmented with the standard
deviation of returns as an additional explanatory variable turned out to be
ambiguous. The authors suspected that the explanatory power of “own”
variance may be spurious due to the high serial correlation of squared
returns. Therefore, to eliminate possible spuriousness, the authors used days
1, 7, 13, ...for estimated returns, observations 3, 9, 15, ...for loadings, and 5,
11, 17, ...for standard deviations. This approach reduced the sample size for
estimation from the initial 2,619 observations down to 436. With these
changes, in 9-out-of-42 groups of stocks, the “own” variance was
statistically significant. Upon implementing the Fama–MacBeth procedure
with the above sampling conventions, just 3-out-of-42 groups had
significant “own” variance. Hence, consistent with the APT, these results
suggested that own variance does not explain expected asset returns.
Another study by Chen (1983) conducted a similar study of the APT by
making comparisons to the CAPM. Using daily returns in the period 1968
to 1978, they used factor analysis to identify five asset pricing factors in
U.S. stock returns. Cross-sectional tests were made using these factors’ beta
loadings. Breaking their sample period into four subperiods, they found that
the five APT factors were significantly priced. In other tests, the CAPM
appeared to be misspecified, as mispricing was captured by the APT. Also,
own variance and size did not improve upon the factor loadings based on
the APT. Consistent with Roll and Ross, based on these and other findings,
he concluded that the evidence supported the APT over the CAPM.
While the empirical evidence supports the notion of multiple factors in
asset pricing models in line with the APT, the identification of the macro
variables associated with systematic risks is problematic. What variables
should be used? Is there precedent in the literature for using specific
variables? Taken together, could the multiple factors in the APT be proxies
for the unobservable theoretical market portfolio in the CAPM? Due to
these questions, the APT remains controversial.

7.3 Summary
Ross (1971, 1974, 1976) developed the arbitrage pricing theory (APT) as a
new general equilibrium framework grounded in a linear relation between
expected returns on assets and their risks. A no-arbitrage condition is
assumed in which no wealth is required due to long/short portfolios held by
investors. Investors earn zero returns for zero risk in well-diversified
portfolios with no asset-specific idiosyncratic risk. Ross showed that the
CAPM is a special case of the APT with a single market factor. More
generally, the APT can have multiple orthogonal factors that investors use
to arbitrage away any excess risk-adjusted return in the market. However,
Ross did not specify the number of factors or their identity in the APT.
To empirically test the APT, Roll and Ross (1980) used the statistical
method of factor analysis in combination with cross-sectional
regression analyses to show that at least three factors exist in U.S. stock
returns. Further tests of own variance as a factor were insignificant in line
with the APT. Another factor analysis study by Chen (1983) found five
significant factors in cross-sectional tests. The empirical findings of these
studies supported the APT.
In sum, the APT provides a general framework for linear factor pricing
models. Researchers building multifactor models in forthcoming chapters
often cite the APT as a theoretical justification for their specifications. Due
to their widespread acceptance among both academic researchers and
investment professionals, multifactor models have become popular over the
past 30 years.

Questions
1. In the CAPM, one of the assumptions is that investors have common
(homogeneous) beliefs about the return distribution of assets. What is
the common assumption about the asset return generating process in
the APT theory?
2.
What are the major strengths and weaknesses of the APT in relation to
the CAPM?
3.
In the single factor case, given that
$$R_i = E(R_i) + \beta _i F + \epsilon _i$$, the APT model specified
as $$E(R_i) = \gamma _0 + \gamma _i\beta$$ is called the exact
APT model by Huberman and Wang (2017). Huberman (1982)
considers the representation
$$E(R_i) = \gamma _0 + \gamma _1\beta _i + \alpha _i$$, which
Huberman and Wang refer to as an approximate APT model. Here it is
assumed that $$\sum _{i = 1}^n\alpha _i = 0$$ and
$$\sum _{i = 1}^n\alpha _i\beta _i = 0$$. How can $$\alpha _i$$s
be interpreted in economic terms?

Problems
1. Consider the following APT model:
$$\begin{aligned} E(R_i) - R_f = \gamma _1\beta _{i,1} +
(7.15)
\gamma _2\beta _{i,2}. \end{aligned}$$
Suppose that the riskless rate $$R_f = 5\%$$, and we have the
following data for three stocks.

Stock i $$E(R_i)$$ $$\beta _{i1}$$ $$\beta _{i2}$$

1 10.1 1.5 0.4
2 8.9 1.2 0.2
3 11.0 1.0

(a)
What are the risk premiums of the factors, i.e., $$\gamma _1$$
and $$\gamma _2$$?
(b)
What are the expected values of the factors?
(c)
What is the stock risk premium for each stock?
 (d) What is $$\beta _{3,2}$$?

2.
Form the equal-weighted portfolio of the stocks in problem 1.
(a)
Compute $$\beta _{i,1}$$ and $$\beta _{i,2}$$ for the
portfolio.
(b)
Compute the expected return of the portfolio.
(c)
Compute the portfolio’s risk premium.
3.
Suppose the expected market return $$E(R_M)$$ equals $$8\%$$.
(a)
Compute CAPM betas for the stocks in problem 1.
(b)
Are the data in problem 1 consistent with both the CAPM and the
APT?

References
Chen, N.-F. 1983. Empirical tests of the theory of arbitrage pricing. Journal of Finance 38: 1393–
1414.
[Crossref]

Connor, G., and R. Korajzyck. 1985. Risk and return in equilibrium APT: Theory and tests. Banking
Research Center Working paper 129, Northwestern University.

Connor, G., and R. Korjzyck. 1986. Performance measurement with the arbitrage pricing theory.
Journal of Financial Economics 15: 373–394.
[Crossref]

Dhrymes, P., I. Friend, and N.B. Gultekin. 1984. A critical reexamination of the empirical evidence
on the arbitrage pricing theory. Journal of Finance 39: 703–738.
[Crossref]

Dybvig, P., and S.A. Ross. 1985. Yes, the APT is testable. Journal of Finance 40: 1173–1188.
[Crossref]
Fama, E.F., and J.D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy 38: 607–636.
[Crossref]

Huberman, G. 1982. A simple approach to arbitrage pricing theory. Journal of Economic Theory 28:
183–191.
[Crossref]

Huberman, G., and Z. Wang. 2017. Arbitrage pricing theory. In The New Palgrave Dictionary of
Economics, ed. M. Vernengo, E.P. Caldentey, and B.J. Rosser Jr. London, UK: Palgrave Macmillan.

Ingersoll, J.E. 1984. Some results in the theory of arbitrage pricing. Journal of Finance 39: 1021–
1039.
[Crossref]

Miller, M.H., and M. Scholes. 1972. Rates of return in relation to risk: A re-examination of recent
findings. In Studies in the Theory of Capital Markets, ed. M.C. Jensen. New York, NY: Praeger.

Roll, R., and S.A. Ross. 1980. An empirical investigation of the arbitrage pricing theory. Journal of
Finance 35: 1073–1103.
[Crossref]

Ross, S.A. 1971. The general validity of the mean variance approach in large markets. Discussion
paper no. 12–72. Rodney L, White Center for Financial Research, University of Pennsylvania.

Ross, S.A. 1974. Return, risk, and arbitrage. In Risk and Return in Finance, ed. I. Friend and J.
Bicksler. New York, NY: Heath Lexington.

Ross, S.A. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13: 341–
360.
[Crossref]

Shanken, J. 1982. The arbitrage pricing theory: Is it testable? Journal of Finance 37: 1129–1140.
[Crossref]

Footnotes
1 See also Huberman (1982).

2 See also Ingersoll (1984) for a generalization of this approach.

3 For example, see Miller and Scholes (1972).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_8

8. Multifactor Models
James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_8

In 1990 William Sharpe was awarded the Nobel Prize in Economics for the
CAPM along with Harry Markowitz for portfolio diversification and
Merton Miller for corporate valuation.1 The Swedish industrialist Alfred
Nobel (inventor of dynamite) dedicated a foundation to recognize
discoveries that benefit humankind. The gold standard for intellectual
achievement, the Nobel Prize is conferred on a person after the true
significance of their discoveries, research, and writings are known. The
CAPM and mean-variance portfolio diversification had become the
centerpieces of modern finance.
Unfortunately, the leading role of the CAPM in asset pricing was cut
short by soon-to-be-published studies. In 1992 Fama and French published
the first in a series of papers that showed the CAPM did not work in the real
world. They documented extensive evidence that there was no relationship
between beta risk and average U.S. stock returns over long periods of time.
Upon reviewing their accumulating evidence in a number of papers, they
concluded that the CAPM was dead. In its place, they proposed a three-
factor model comprised of the market factor plus two factors based on the
firm characteristics of size (i.e., small and large firms) and book
equity/market equity ratios (i.e., value and growth firms). Their empirical
tests showed that this new multifactor model substantially improved the
goodness-of-fit of the CAPM to stock returns. Not surprisingly, some
degree of controversy was sparked by the three-factor model. In this chapter
we review the three-factor model and resultant controversy in the asset
pricing literature.

8.1 What Explains Stock Returns?

If CAPM beta does not explain stock returns very well, what does? Fama
and French (1992) set the stage for their later multifactor models by
showing that two simple firm characteristics—size and book-to-market
equity—can help to explain the cross-section of average stocks returns in
the United States. They cited earlier research to support their findings. The
size effect on stock returns goes back to Banz (1981), who found that small
(large) firms had average returns that were too high (low) based on their
betas. Statman (1980) and Rosenberg et al. (1985) found that stock returns
were related to the ratio of book equity to market equity (denoted hereafter
as BM). Other researchers detected relations between stock returns and
other firm characteristics, including financial leverage and earnings/price
ratios.2
To prove the relation of stock returns to size and BM characteristics of
firms, Fama and French collected U.S. stock returns from 1963 to 1990.
Stocks were formed into 10 deciles in terms of size (as measured by market
capitalization equal to stock price times number of shares outstanding) as
well as BM. Also, they formed decile portfolios of stocks using estimated
market model betas. In the sample period, average betas ranged from 0.87
for the low decile to 1.72 for the high decile. This range of beta values is
representative of the population of U.S. stocks in their sample period. The
corresponding average returns for these low and high beta deciles were
1.34% per month and 1.14%, respectively, which is opposite of CAPM
theory with expected returns increasing as beta increases.
While beta did not appear to explain stock returns in a reliable manner,
size and BM did. Table 8.1 documents selected results reported by Fama
and French (1992, Table V, p. 446). Looking at the columns in the table, at
any given BM ratio, with the exception of the low decile (high growth)
stocks in the column on the left, average returns per month (in percent) tend
to decrease as size increases from the low decile (small stocks) to the high
decile (big stocks). Also, referring to the rows, for any given size, average
returns tend to increase as BM increases from the low decile (growth
stocks) to the high decile (value stocks). The highest earnings stocks are
small, growth stocks in the upper right-hand corner of the table (e.g., a
maximum average return of 1.92% for stocks that are in the low decile by
size and high decile by BM). Again, high growth stocks in the low BM
decile do not follow the return pattern of other stocks. Nonetheless, the size
and BM returns patterns in Table 8.1 are quite convincing. Simple sorts on
size and BM firm-level characteristics tell the story—stock returns are
related to size and BM.
Table 8.1 Average monthly returns for U.S. stock portfolios sorted by size and book-to-market
equity (BM) from July 1963 to December 1990 in the Fama and French (1992) study

Book-to-market equity (BM)

Size Low decile Decile 3 Decile 5 Decile 7 High decile
Small decile 0.70 1.20 1.56 1.70 1.92
Decile 3 0.56 1.23 1.36 1.30 1.60
Decile 5 0.88 1.08 1.13 1.44 1.49
Decile 7 0.95 0.99 0.99 0.99 1.47
Big decile 0.89 0.84 0.79 0.81 1.18

This table reports the average monthly returns (in percent) for U.S. stock
portfolios sorted into deciles by size and book-to-market equity (BM). Size
is measured by the market capitalization of stocks (i.e., stock price times
shares outstanding). The sample period is July 1963 to December 1990. The
selected results shown below are based on Fama and French (1992, Table V,
p. 446)
 Next, the authors ran formal statistical tests using Fama and
MacBeth (1973) cross-sectional regressions using different models. The
time-series models contained a number of variables, including the market
factor, size, BM, leverage, and the earnings/price ratio. The market price of
risk for CAPM beta ( $$\lambda _m$$) was insignificant. However,
the market prices of beta coefficients associated with size and BM (or
loadings) were very significant. Their market prices were −0.15% and
0.50% per month with significant t-statistics equal to −2.58 and 5.71,
respectively. These results confirmed that bigger stocks have lower returns
than smaller stocks (i.e., the negative price of risk), and that high BM
(value) stocks have higher returns than lower BM (growth) stocks (i.e.,
positive price of risk). Also, loadings for leverage and the earnings/price
ratio were not significant when size and BM beta loadings were included in
the model. Therefore, they inferred that these firm characteristics were
absorbed (or accounted for) by the size and BM variables. From these
findings, the authors concluded that size and BM beta risk loadings help to
explain stock returns but not CAPM beta. In their final remarks, concerning
the CAPM, they commented:

Black et al. (1972) and Fama and MacBeth (1973) find that, as
predicted by the model, there is a positive simple relation between
average return and market $$\beta$$ during the years (1926-1968)
... we find that this simple relation between $$\beta$$ and average
return disappears during the more recent 1963-1990 period. (Fama
and French 1992, p. 449)

This sensitivity of the CAPM to the sample period was observed by

Jagannathan and McGrattan (1995) also. In the U.S. stock market, they
found a noticeable relation between average returns and beta in the long
sample periods of 1926 to 1975 and 1926 to 1991 but not in the shorter
periods 1976 to 1980 and 1981 to 1991. In the latter periods, size and BM
factors did better than the market factor at explaining cross-sectional
variation in average stock returns. Thus, the CAPM does not consistently
hold over time, which implies that it should be supplemented with other
factors.
The 1992 Fama and French study is an important landmark in the asset
pricing literature because it set the stage for multifactors that better fit stock
return data. One year later, Fama and French would propose their seminal
multifactor model.

8.2 Fama and French Three-Factor Model

Akin to the APT of Ross (1976) and Fama and French (1993) utilized size
and BM variables to construct long/short portfolios as factors. These
long/short or zero-investment portfolios are also called mimicking
portfolios. The betas or loadings on these factors measure the sensitivity of
asset returns to the factors. They sort stocks into six portfolios in a 2
$$\times$$3 matrix by size and value (BM) as shown in Fig. 8.1. The size
factor is defined as the average of the three small (blue) portfolios’ returns
minus the three big (red) portfolios’ returns. The value factor is the average
of the two value (green) portfolios’ returns minus the two growth (brown)
portfolios’ returns. They used more value than size portfolios due to the
stronger relation of BM to stock returns than size in their 1992 paper. They
denoted the size factor as SMB for small minus big and the value factor as
HML for high minus low BM. Both factors were computed using value-
weighted (not equal-weighted) returns on a monthly basis.

Fig. 8.1 Stocks are sorted into a 2 $$\times$$ 3 matrix by size and value (BM). The size factor is
the average returns of the three small (blue) portfolios minus the three (red) big portfolios, and the
value factor is average returns of the two value (green) portfolios minus the two growth (brown)
portfolios

Together with the market factor, they proposed a three-factor

model with the added multifactors of size and value. The market factor was
the value-weighted excess return computed as the average CRSP stocks’
return minus the Treasury bill rate each month. The CRSP market index
contains all U.S. stocks in the database created by the Center for Research
in Security Prices at the University of Chicago. The three-factor model is:
$$\begin{aligned} R_{it}-R_{ft}=\alpha _i+\beta _{i,m} (R_{mt}-
R_{ft})+\beta _{i,S} SMB _t+\beta _{i,V} HML _t+e_{it}, (8.1)
\end{aligned}$$
where $$\beta _{i,S}$$ and $$\beta _{i,V}$$ estimate the
sensitivity of excess returns of the ith stock or stock portfolio to the size and
value factors, respectively, and other notation is as in previous chapters.
Figure 8.2 gives a four-dimensional visualization of the model in risk/return
space. Each axis in this figure is perpendicular (or orthogonal) to all the
other coordinate axes.3

Fig. 8.2 The Fama and French three-factor model with excess returns a function of market, size, and
value factors takes into account four dimensions in return/risk space

To estimate the three-factor model, dependent variable excess returns on

the left-hand-side of Eq. (8.1) were based on 25 portfolios sorted on size
and BM (similar to the six size/BM portfolios in their factors). Value-
weighted returns were again computed on a monthly basis. These portfolio
returns are the test assets. In the sample period 1963 to 1991, they found
that the market risk premium was 0.43% per month compared to the size
risk premium (or SMB return) of 0.27% and the value risk premium (or
HML return) of 0.40%. Hence, the value factor was larger in magnitude
than the size factor.
Fama and French estimated the three-factor model using monthly U.S.
stock returns in the period July 1963 to December 1991. The results
strongly supported their model. The beta coefficients denoted
$$\beta _{i,m}$$, $$\beta _{i,S}$$, and $$\beta _{i,V}$$
corresponding to the market, size, and value factors, respectively, were all
highly statistically significant with very high t-statistics in most of the 25
size and BM portfolios. The magnitudes of the beta coefficients followed
definite patterns: (1) small stocks had larger coefficients
$$\beta _{i,S}$$ related to the size factor SMB than big stocks; and
(2) high BM (value) portfolios had larger coefficients
$$\beta _{i,V}$$ related to value factor HML than low BM (growth)
portfolios. These patterns suggest that small (value) stocks have relatively
higher returns than big (growth) stocks on average. It is important to
mention that the size and value factors were uncorrelated with one another
for the most part, but both had some correlation with the market factor. The
latter is not surprising as the size and value portfolios are comprised of
stocks that represent large components of the market portfolio. Hence, there
could be some amount of multicollinearity in the three-factor model, which
can lead to unstable regression parameter estimates as discussed in Chapter
4.
To measure the goodness-of-fit of different models, $$R^2$$
estimates were computed. Table 8.2 summarizes their findings. In Panel A
of this table, we see that the CAPM has fairly strong goodness-of-fit with
$$R^2$$ values in the range of 61% to 92%. There is a tendency for
$$R^2$$ values to be higher for low BM (growth) and big size stocks
than other stocks. Looking at Panel B with three-factor model results, the
$$R^2$$ values are relatively higher in the range of 79% to 97%.
Indeed, 20-out-of-25 $$R^2$$ values exceed 90%, and 10-out-of-25
are equal to or greater than 95%. The relatively high and in many cases near
perfect goodness-of-fit of these size and BM sorted portfolios suggests that
the model is empirically valid with respect to the 25 size and BM test asset
portfolios.
Table 8.2 Goodness-of-fit $$R^2$$ estimates for 25 U.S. stock portfolios sorted on size and
book-to-market equity (BM) from July 1963 to December 1991 in the Fama and French (1993) study

Low BM Quintile 2 Quintile 3 Quintile 4 High BM

Panel A. CAPM $$R^2$$ estimates
Small size quintile 0.67 0.70 0.68 0.65 0.61
Quantile 2 0.79 0.79 0.76 0.76 0.71
Quantile 3 0.84 0.84 0.80 0.79 0.74
Quantile 4 0.89 0.90 0.87 0.80 0.76
Big size quintile 0.89 0.92 0.84 0.79 0.69
Panel B. Three-factor model $$R^2$$ estimates
Small size quintile 0.94 0.96 0.97 0.97 0.96
Quantile 2 0.95 0.96 0.95 0.95 0.96
Quantile 3 0.95 0.94 0.93 0.93 0.93
Quantile 4 0.94 0.93 0.91 0.89 0.89
Big size quintile 0.94 0.88 0.90 0.79 0.83

This table reports the $$R^2$$ estimates for the CAPM and Fama and
French three-factor model using U.S. stock portfolios sorted into quintiles
by size and book-to-market equity (BM). Size is measured by the market
capitalization of stocks (i.e., stock price times shares outstanding). The
sample period is July 1963 to December 1991. The selected results shown
below are based on Fama and French (1993, Tables 4 and 6, pp. 20 and 24)

They also tested a bond market model using government bond

portfolios with maturities from 1-to-10 years and five corporate bond
portfolios with different Moody’s credit ratings. A term-to-maturity risk
factor denoted TERM was defined as the long-term bond return minus one-
month Treasury bill rate. A default risk factor denoted DEF was the
differenced returns of long-term corporate bonds minus long-term
government bonds. As we have seen, the three-factor model performed well
for the 25 stock portfolios, but this feat was not repeated for the bond
portfolios. Likewise, the bond factors were significant for bond portfolios
but not stock portfolios. Since the bond factors tended to be explained by
the market factor, once the market factor was added to the stock model,
there was little support for the bond factors. Of course, the market factor
contains the Treasury bill rate, which likely explains the weak effect of
bond factors.
To cross-sectionally test their three-factor model, the intercepts (
$$\alpha$$s) for different models were estimated for the 25 size and
BM portfolios. Citing Merton’s (1973) ICAPM paper, Fama and French
argued that a well-specified model should have intercepts equal to zero (i.e.,
no mispricing error). Table 8.3 summarizes some of their $$\alpha$$
test results.
Table 8.3 Intercept or $$\alpha$$ estimates for 25 U.S. stock portfolios sorted on size and book-
to-market equity (BM) from July 1963 to December 1991 in the Fama and French (1993) study

Low BM Quintile 2 Quintile 3 Quintile 4 High BM

Panel A. CAPM $$\alpha$$ estimates
Small size quintile 0.31 0.62* 0.71** 0.80*** 0.92***
Quintile 2 0.35 0.63* 0.77*** 0.75*** 0.93***
Quintile 3 0.34 0.58** 0.60** 0.73*** 0.89***
Quintile 4 0.41 0.27 0.49** 0.69*** 0.96***
Big size quintile 0.34 0.30 0.25 0.50** 0.53**
Panel B. Three-factor model $$\alpha$$ estimates
Big size quintile −0.34*** −0.12 −0.05 0.01 0.00
Quintile 2 −0.11 −0.01 0.08 0.03 0.02
 Low BM Quintile 2 Quintile 3 Quintile 4 High BM
Quintile 3 −0.11 0.04 −0.04 0.05 0.05
Quintile 4 0.09 −0.22*** −0.08 0.03 0.13
Big size quintile 0.21*** −0.05 −0.05 −0.05 −0.17

This table reports the monthly $$\alpha$$ estimates (in percent) for the
CAPM and Fama and French three-factor model using U.S. stock portfolios
sorted into quintiles by size and book-to-market equity (BM). Size is
measured by the market capitalization of stocks (i.e., stock price times
shares outstanding). The sample period is July 1963 to December 1991. The
selected results shown below are based on Fama and French (1993, Table
9a, pp. 36–37)
Note Asterisks indicate that the $$\alpha$$ estimate is statistically
different from zero at the *−10%, **−5%, and ***−1% levels

In Panel A of Table 8.3, we see the results for the CAPM. The
$$\alpha$$ estimates for the portfolios using the market factor alone
produced positive and significant intercepts. For example, the small size
quintile and high BM quintile portfolio has a monthly $$\alpha$$
equal to 0.92%, which is highly significant at the 1% level and therefore
different from zero. This $$\alpha$$ corresponds to a very large
mispricing error of about 11% per year. Except for low BM (growth) stocks,
the $$\alpha$$ estimates for most portfolios are significantly greater
than zero. When they repeated these tests using only the SMB and HML
factors in a two-factor model, the $$\alpha$$ estimates were again
very positive and significant with a high of 0.79% per month, or about 9.5%
per year.
In stark contrast to the CAPM and two-factor model with size and value
factors, as shown in Panel B of Table 8.3, upon combining the market factor
with SMB and HML factors in the three-factor model, the intercepts became
close to zero. With the exception of only 3-out-of-25 portfolios, the
intercepts were not significantly different from zero. These results strongly
support the three-factor model.
The bond factors TERM and DEF did not further reduce the intercepts
for stock portfolios but did so for bond portfolios. Based on these and other
tests, Fama and French (1993, p. 41) inferred that “... the three-factor
model does a good job on the cross-section of average stock returns.” They
interpreted the size and value factors as proxies for risk factors associated
with these firm-level characteristics (akin to the CRSP excess return serving
as a proxy for the true market factor in the CAPM).
In their paper Fama and French conjectured that Merton’s (1973)
ICAPM and Ross’ (1976) APT provided some theoretical support for their
three-factor model. Recall that these theoretical models proposed multiple
risk factors to price assets in equilibrium. In this regard, the three-factor
model is mainly justified on empirical grounds—that is, it noticeably
improves the goodness-of-fit to stock return data compared to the CAPM.
In their paper entitled “The CAPM is Wanted, Dead or Alive,” Fama and
French (1996) argued that the failures of the CAPM to explain expected
returns motivate the investigation of return anomalies which can be taken
into account by multifactor ICAPM and APT models. In their view, the
CAPM was dead.
Summarizing their new long/short zero-investment portfolio,
multifactor model approach to asset pricing, Fama and French were mindful
of the questions left unanswered. In their words,

... our work leaves many open questions. Most glaring, we have not
shown how the size and book-to-market factors in returns are driven
by the stochastic behavior of earnings ... Can specific fundamentals
be identified as variables that lead to common variation in returns
that is independent of the market and carries a different premium
than general market risk? These and other interesting questions are
left for future work. (Fama and French 1993, p. 55)

Returning to the above question of the role of earnings, Fama and French
(1996) published another paper in which they considered the following dual
hypotheses: (1) size and value factors are related to some common factor;
and (2) size and value patterns are related to the behavior of earnings which
affects stock prices. They contended that high BM (value) ratios with low
relative stock prices signal low earnings and possible financial distress, and
vice versa for low BM (growth) ratios with high relative stock prices.
Moreover, holding BM constant, small companies tend to have lower
earnings than big companies. While it is possible that small companies’
earnings fluctuate with the business cycle and exceed those of big
companies at times, in the long run small companies underperform big
companies in terms of profits over time.
To investigate these dual research hypotheses, Fama and French
collected data on size, BM, earnings, and stock returns for U.S. companies
in the period 1963 to 1992. Using simple descriptive analyses, they
documented that size and BM factors based on companies’ earnings were
similar to those computed using stock returns. In other words, common
factors in earnings and returns were found. However, while market and size
factors in earnings were linked to the market and size factors in returns, the
BM factor in returns did not appear to be closely associated with the BM
factors in earnings. Contrary to their hypothesis, equity returns were not
related to the BM factor in earnings.
In sum, there is some evidence to support the notion that common
variations in returns are related to common factors in earnings. As such,
they concluded by posing the following questions for future research:

(i) What are the underlying state variables that produce variation in
earnings and returns related to size and BE/ME?
(ii) Do these unnamed state variables produce variation in
consumption and wealth that is not captured by the overall market
factor and so can explain the risk premiums associated with size and
BE/ME? (Fama and French 1996, p. 154)

Unfortunately, given that state variables are proxied by macro level

information (e.g., gross national product (GNP), aggregate wealth and
consumption, employment, etc.), they were doubtful about the prospects of
finding state variable determinants or drivers of earnings and returns. This
pessimistic assessment of finding empirical models in the true spirit of the
ICAPM and APT laid down a clear challenge to the asset pricing world. A
challenge that many researchers accepted as we will see in the next chapter.

8.3 Three-Factor Controversy

Black (1993, p. 75) later criticized the three-factor model on a number of
grounds. First, Black complained of potential data mining. This problem is
a result of researchers exploring many different independent variables on
the right-hand-side of the time-series regression equation, different time
periods, and alternative combinations of variables in different models.
Eventually, if you run enough data, you will find something that looks
statistically significant and then report those results to the exclusion of all
the other insignificant results. This kind of data snooping could produce
results that are entirely accidental or random. He singled out anomalies in
finance literature as a classic example of this problem. The size effect in the
Fama and French (1992) study was an anomaly that disappeared in the 1981
to 1990 period with little or no impact on stock returns. Also, there is no
clear explanation for why size and expected returns or BM ratios and
expected returns are related to one another. For these reasons, he
commented: “Lack of theory is a tipoff: watch out for data mining!” Black
concluded that it was premature to claim the death of beta and the CAPM.
In a short paper, Black (1995) observed that theory posits at least one
priced factor. Everyone agrees (including Fama and French) that the
expected excess return in the market as a whole is positive. He reasoned
that, if the relation between expected returns and beta was flat, then the
expected return on the second factor related to the zero-beta portfolio in his
zero-beta CAPM must be large. Interestingly, due to the fact that size and
BM ratios sometimes are not related to expected returns and lack theoretical
foundations as discussed above, he referred to the size and value factors as
third and fourth factors, respectively. He believed these multifactors were of
lesser importance than his zero-beta factor.4
An often-cited study by Kothari et al. (1995) published findings that
contradicted Fama and French. Using annual (not monthly) U.S. stock
returns for the longer sample period 1927 to 1990, they found that the
annual contribution of beta risk was 8.9% to 11.7% for the equally weighted
stock market index and 6.2% to 8.9% for the value-weighted stock market
index. Note here that the equal-weighted index weights small stocks more
heavily, as the value-weighted index weights them less than big stocks. So a
size effect is apparent, with small stocks having higher returns than big
stocks. Time-series regression was used to estimate betas over the sample
period, and then the market price of beta was estimated using in-sample
cross-sectional regression. This two-step Fama and MacBeth procedure was
used on both a monthly and annual basis by rolling forward one month and
one year at a time. Cross-sectional regressions were estimated using both
monthly and annual returns as dependent variables and betas estimated
using annual returns for the entire sample period as the independent
variable. The average of these estimated $$\lambda _m$$s over the
entire sample period had t-statistics greater than 2.50 in a number of
different test asset portfolios. Thus, beta risk loadings were priced in the
cross-section of average stock returns for annual betas. They repeated these
analyses in the subperiod 1941 to 1990 with similar results. Thus, based on
annual returns to estimate market beta instead of monthly returns, their
empirical tests supported the CAPM.
Using industry portfolios as test assets, the value factor was only
weakly related to average stock returns. In this regard, the stocks used in the
Fama and French tests were suspected of survivor bias as firms in financial
distress were dropped from their analyses. Value firms with high book-to-
market equity ratios were less likely to survive over time and therefore be
included in sample data. Consequently, they argued that Fama and French
results refuting the CAPM were subject to some degree of sample period
selection as well as sample stock selection. Betas estimated using annual
(not monthly) returns and broader samples including industry portfolios
appeared to better support the CAPM and somewhat diminish support for
the size and value factors.5 Kothari et al. (1995, p. 221) concluded:

Finally, we emphasize that the failure of a significant relation

between B/M and return to emerge for the S &P industry portfolios
... poses a serious challenge to the B/M “empirical asset pricing
model” ... A useful pricing model must be trusted to work under a
wide variety of conditions and not just for a limited set of portfolios.

Echoing this concern, researchers have contended that an endogeneity

problem arises when the test assets used to test the size and value factors are
themselves based on size and value firm-level characteristics. In the
extreme, if you regress Y on Y, a perfect goodness-of-fit will be achieved.6
Similar to Kothari, Shanken, and Sloan , a number of authors, including
Lewellen et al. (2010), Daniel and Titman (2012), and others, have
recommended that the test assets should be expanded to include portfolios
that are less strongly correlated with proposed factors. Industry
portfolios are considered to be a good choice due to being exogenous (or
external) to various factors for the most part.
It was not long before Fama and French (1996) responded to these
criticisms. They countered that: (1) beta is insufficient to fully explain
average stock returns; and (2) size and BM are uncorrelated to beta but add
more significantly to explaining returns than beta. Even if beta is significant
in some instances, empirical tests suggested that multifactor forms of the
ICAPM and APT were superior to the CAPM. It is important to recognize
that Fama and French did not say that the market factor is useless; rather,
the market factor must be augmented with other factors to more fully
account for average stock returns. Their main argument was that the CAPM
does not hold—multifactor models are needed in asset pricing. Consistent
with this argument, in another study of 12 different countries plus the
United States, Fama and French (1998) presented corroborating evidence
that the three-factor model outperformed the CAPM.
As recognized by Fama and French, it is possible that the CAPM fails
due to poor (inefficient) proxies for the true (efficient) market portfolio. In
Chapter 4 we cited Roll’s (1977) critique that the CAPM cannot be tested
without an efficient empirical proxy m for the theoretical market portfolio
M. In the words of Fama and French (1996, p. 1957): “... valid applications
of the CAPM await the coming of M.” They later commented in a 2004
survey article on the CAPM that “... If betas do not suffice to explain
expected returns, the market portfolio is not efficient, and the CAPM is
dead in its tracks” (Fama and French 2004, p. 36). In other words, even if
the market premium is significant, other factors such as size and value are
needed to more fully explain expected returns.
Fama and French (1996) inferred from their empirical evidence that
either multifactor ICAPM and APT models explain the failure of the CAPM
or investors are irrational. The former possibility arises from the reasonable
inference that investors are concerned with more than beta risk associated
with the market factor. Risk-averse investors could plausibly care about a
number of different risks that exist in the market. The latter possibility
relates to the behavioral school for explaining asset pricing anomalies such
as size, value, etc. that the CAPM failed to explain (see Chapter 2). For
example, investors price growth firms with low BM ratios at too high a
stock price that results in low returns; conversely, value firms with high BM
values are distressed and therefore priced too low by investors that result in
high returns. If indeed irrational market behavior based on emotional
investor responses is important in financial markets, then the efficient
markets assumption of asset pricing models is weakened. In either (or both)
case(s), asset pricing is more complex than originally conceived by
Markowitz, Sharpe, Black, Merton, Ross, and many others.
In search of better models, researchers began to explore long/short zero-
investment portfolios to see if other significant factors might exist. The hunt
was on—the search for anomalies in stock returns accelerated in the years
to come in efforts to identify new factors. Forthcoming Chapter 9 reviews
some of the most prominent of these factors and models.

8.4 Summary
After extensive tests using U.S. stock returns over many years, Fama and
French (1996) declared the CAPM dead. Little or no relation between U.S.
stock returns and beta risk led them to propose a new multifactor model. In
their papers, Fama and French (1992, 1993, 1995, 1996) introduced a three-
factor model, which proved to boost the empirical fit to stock returns of the
CAPM’s market factor by using the firm characteristics of size and value
(i.e., book-to-market equity ratios denoted BM). A major innovation
consistent with Ross’ (1976) APT was the construction of long/short zero-
investment factors—for example, long small stocks and short big stocks as
well as long high BM stocks and short low BM stocks. These so-called size
and value factors in combination with the market factor substantially
lowered intercept ( $$\alpha$$) estimates that reflect mispricing in the
time-series regression estimation of asset pricing models. Related loosely to
the theoretical ICAPM and APT models of Merton (1973) and Ross (1976),
respectively, the three-factor model is foundational in the evolution of asset
pricing. In an effort to link their three-factor model to economic
fundamentals, they showed that the size and value factors in stock returns
have some relationship to size and value factors in firm earnings. However,
the identification of ICAPM and APT state variables that drive the linkage
between size plus BM firm characteristics and returns/earnings was left an
unanswered question for future research.
Controversy surrounding the three-factor model started with
Black’s (1993) criticism that it was a product of data mining rather than
theory. While it improved the empirical fit of the CAPM to stock return
data, the state variables underlying the size and value factors were
unknown. Also, Black (1995) reasoned that, given positive expected excess
returns in the market over time, if the return/beta relation is flat, then the
zero-beta factor from his zero-beta CAPM must be large to explain market
returns. The size and value factors were not always present in different
sample periods and, more importantly, lacked theory. A well-known study
by Kothari et al. (1995) found evidence that annual stock returns over long
periods of time were consistent with a linear relation between market beta
and the cross-section of average stock returns. Fama and French (1996)
countered that the CAPM failed in some periods but worked in others; in
both cases, the size and value factors boosted the explanatory power of the
CAPM. Kothari et al. also argued that the value factor did not work well
with industry portfolios and that survivor bias exists among high book-to-
market equity ratio firms that are distressed and can drop out of data
samples in some time periods. Relatedly, some researchers have cited an
endogeneity problem wherein the test asset portfolios’ returns (i.e., the
dependent variable in an asset pricing model) are closely related by
construction to the factors (i.e, the independent variables in a model).
Naturally the model would fit this data quite well. Hence, Lewellen et al.
(2010), Daniel and Titman (2012), and others have recommended that other
test assets portfolios exogenous to the factors should be used to investigate
the validity of an asset pricing model. Industry portfolios were believed to
be good exogenous portfolios.
Fama and French (1996, 2004) defended the three-factor model due to
the failure of the CAPM. According to their logic, the market portfolio is
not efficient which means the CAPM is dead. Two possibilities help to
explain this failure: (1) other factors exist as theorized by the ICAPM and
APT; and (2) investors can be irrational at times that causes anomalies in
asset returns unrelated to risk as proposed by the behavioral school. The
ensuing search for multifactors and anomalies in stock returns triggered a
revolution in asset pricing that continues today. As we will see in
forthcoming Chapter 9, many new long/short factors have been proposed by
researchers that go well beyond the prototypical three-factor model.

Questions
1. If CAPM beta does not explain stock returns, what does? In this
regard, Fama and French (1992) found that stock returns were related
to two firm characteristics. What are these characteristics and briefly
discuss what they found.
2.
Fama and French (1993) used the size and BM variables to construct
zero-investment factors. Define these factors and write the equation
for their resultant three-factor model. What are the multifactors in this
model?
3.
Fama and French (1993) conducted tests of their proposed three-factor
model. What were the test asset portfolios? What did they find?
Discuss the results for estimated beta coefficients and goodness-of-fit
of their model.
4.
How did Fama and French (1993) cross-sectionally test the three-
factor model? What did they find in test results?
5.
Fama and French (1993) specified and tested a bond model. What did
their results show?
6.
If the CAPM is dead, and asset prices are better explained by multiple
factors, what theoretical support for multifactor models exists?
7.
Fama and French (1996) argued that size and value factors are linked
to economic fundamentals such as earnings within the firm. What
were their arguments in this regard?
8.
Did Fama and French (1996) find evidence to support a link between
earnings and the size and value factors? Review their findings. What
are the underlying state variables that produce variation in earnings
and stock returns related to size and BM ratios?
9.
Controversy surrounded the three-factor model due to skepticism by
researchers. Black (1993, 1995) said that the model had two problems.
Discuss these two potential problems.
10.
Kothari, Shanken, and Sloan (1995) found evidence to support the
CAPM. Also, they criticized the Fama and French studies. Review
these issues in their study.
11. What is a possible endogeneity problem in the Fama and French asset
pricing tests of the three-factor model? How can we reduce this
bl i t i i t t?
 problem in asset pricing tests?
12.
How did Fama and French (1996) counter criticisms of their three-
factor model? Do you think that multifactor models are needed to
more fully price assets or are investors irrational leading to anomalous
returns in the market not explained by the CAPM market factor?

Problems
1.
Assume that stocks are sorted into a 2 $$\times$$3 matrix of size and
value (BM). How did Fama and French (1993) use this matrix to create
their size and value factors?
2.
Draw a diagram that depicts Fama and French’s three-factor model.
How many dimensions are there in your drawing?

References
Banz, R.W. 1981. The relationship between return and market value of common stocks. Journal of
Financial Economics 9: 3–18.
[Crossref]

Basu, S. 1977. Investment performance of common stocks in relation to their price-earnings ratios: A
test of the efficient market hypothesis. Journal of Finance 12: 129–156.

Bhandari, L.C. 1988. Debt/equity ratio and expected common stock returns: Empirical evidence.
Journal of Finance 43: 507–528.
[Crossref]

Black, F. 1972. Capital market equilibrium with restricted borrowing. Journal of Business 45: 444–
454.
[Crossref]

Black, F. 1993. Beta and return. Journal of Portfolio Management 20: 8–18.
[Crossref]

Black, F. 1995. Estimating expected return. Financial Analysts Journal 49: 36–38.
[Crossref]

Black, F., M.C. Jensen, and M. Scholes. 1972. The capital asset pricing model: Some empirical tests.
In Studies in the Theory of Capital Markets, ed. M.C. Jensen. New York, NY: Praeger.

Clare, A.D., R. Priestley, and S.H. Thomas. 1998. Reports of beta’s death are premature: Evidence
from the UK. Journal of Banking and Finance 22: 1207–1229.
[Crossref]

Daniel, K., and S. Titman. 2012. Testing factor-model explanations of market anomalies. Critical
Finance Review 1: 103–139.
[Crossref]

Fama, E.F., and K.R. French. 1992. The cross-section of expected stock returns. Journal of Finance
47: 427–465.
[Crossref]

Fama, E.F., and K.R. French. 1993. The cross-section of expected returns. Journal of Financial
Economics 33: 3–56.
[Crossref]

Fama, E.F., and K.R. French. 1995. Size and book-to-market factors in earnings and returns. Journal
of Finance 50: 131–156.
[Crossref]

Fama, E.F., and K.R. French. 1996. The CAPM is wanted, dead or alive. Journal of Finance 51:
1947–1958.
[Crossref]

Fama, E.F., and K.R. French. 1998. Value versus growth: The international evidence. Journal of
Finance 53: 1975–1999.
[Crossref]

Fama, E.F., and K.R. French. 2004. The capital asset pricing model: Theory and evidence. Journal of
Economic Perspectives 18: 25–46.
[Crossref]

Fama, E.F., and J.D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy 81: 607–636.
[Crossref]

Gibbons, M.R., S.A. Ross, and J. Shanken. 1989. A test of the efficiency of a given portfolio.
Econometrica 57: 1121–1152.
[Crossref]

Gordon, M.J. 1962. The Investment, Financing, and Valuation of the Corporation. Homewood, IL:
R.D Irwin.

Hsia, C.-C., B.R. Fuller, and B.Y.J. Chen. 2000. Is beta dead or alive? Journal of Business Finance
and Accounting 27: 283–311.
[Crossref]

Jagannathan, R., and E.R. McGrattan. 1995. The CAPM debate. Quarterly Review 19 (4, Fall): 2–17.
Federal Reserve Bank of Minneapolis.

Kolari, J.W., W. Liu, and J. Huang. 2021. A New Capital Asset Pricing Model: Theory and Evidence.
New York, NY: Palgrave Macmillan.
[Crossref]
Kothari, S.P., J. Shanken, and R.G. Sloan. 1995. Another look at the cross-section of expected stock
returns. Journal of Finance 50: 185–224.
[Crossref]

Lakonishok, J., and A.C. Shapiro. 1986. Systematic risk, total risk, and size as determinants of stock
market returns. Journal of Banking and Finance 10: 115–132.
[Crossref]

Lakonishok, J., and A. Shleifer. 1984. Stock returns, beta, variance and size: An empirical analysis.
Financial Analysts Journal 40: 36–41.
[Crossref]

Lewellen, J., S. Nagel, and J.A. Shanken. 2010. A skeptical appraisal of asset pricing tests. Journal
of Financial Economics 96: 175–194.
[Crossref]

Lintner, J. 1965. The valuation of risk assets and the selection of risky investments in stock portfolios
and capital budgets. Review of Economics and Statistics 47: 13–37.
[Crossref]

Markowitz, H.M. 1959. Portfolio selection: Efficient diversification of investments. New York, NY:
Wiley.

Merton, R.C. 1973. An intertemporal capital asset pricing model. Econometrica 41: 867–887.
[Crossref]

Modigliani, F., and M.H. Miller. 1958. The cost of capital, corporation finance and the theory of
investment. American Economic Review 48: 261–297.

Modigliani, F., and M.H. Miller. 1963. Corporation income taxes and the cost of capital: A
correction. American Economic Review 53: 433–443.

Mossin, J. 1966. Equilibrium in a capital asset market. Econometrica 34: 768–783.

[Crossref]

Roll, R. 1977. A critique of the asset pricing theory’s tests, part I: On past and potential future
testability of the theory. Journal of Financial Economics 4: 129–176.
[Crossref]

Rosenberg, B., K. Reid, and R. Lanstein. 1985. Persuasive evidence of market inefficiency. Journal
of Portfolio Management 11: 9–17.
[Crossref]

Ross, S.A. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13: 341–
360.
[Crossref]

Sharpe, W.F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance 19: 425–442.

Statman, D. 1980. Book values and stock returns. The Chicago MBA: A Journal of Selected Papers
4: 25–45.
Treynor, J. L. 1961. Market value, time, and risk. Unpublished manuscript.

Treynor, J. L. 1962. Toward a theory of market value of risky assets. Unpublished manuscript.

Footnotes
1 See their original works, especially Sharpe (1964), Markowitz (1959), and Modigliani and Miller
(1958, 1963). With respect to the CAPM, both Lintner (1965) and Mossin (1966) had died by 1990
and Treynor (1961, 1962) had not never published his CAPM papers, such that these authors were
not eligible for the prize.

2 See Basu (1977), Lakonishok and Shapiro (1986), and Bhandari (1988). Fama and French (2004)
have provided an excellent survey of the CAPM and these contradictory studies.

3 We cannot visualize four-dimensional objects as humans, but a multivariate regression model

enables a mathematical realization of N dimensions corresponding to N variables including both the
dependent and independent variables. Interestingly, Einstein’s theory of General Relativity argued
that a space-time continuum exists with 3 dimensions of space and 1 dimension of time. He thought
that there is curvature in space-time due to gravity. Could the return-risk space be curved also? Is
there a unifying force like gravity that links returns and risk in financial markets? Perhaps it can bend
return-risk space much as gravity can bend light as it travels through space.

4 In Chapter 10 we cover the ZCAPM of Kolari et al. (2021), which is based on Black’s (1972) zero-
beta CAPM. Consistent with Black’s (1995) comments about a large second factor, empirical tests of
the ZCAPM have revealed a highly significant, large second factor (viz.,cross-sectional returncross-
sectional return dispersion in the market).

5 See also Hsia et al. (2000), who used annual returns to compute moving-average betas and better
take into market anomalies. Their results supported a positive relation between average returns and
beta risk also. Additionally, a study of U.K. stock market returns by Clare et al. (1998) found that the
beta was not dead. Using a somewhat different method of estimating the cross-sectional relation
between expected returns and factor loadings, market beta risk was significantly priced in their tests.
However, their study only employed 100 U.K. stocks with data available from 1980 to 1993. Thus,
their results were somewhat limited in scope.

6 In time-series regressions of the three-factor model, information in the size and value test asset
portfolios (and even residual errors for these portfolios) will naturally have information related to the
size and value factors. In this respect, recall from Chapter 4’s discussion of the market model and
underlying statistical assumptions, the error terms should be uncorrelated with the independent
variables.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_9

9. Anomalies and Multifactor Models

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_9

Fama and French (1992, 1993, 1995, 1996) published a series of papers in
the 1990s that replaced the CAPM with a three-factor model containing size
and value factors. Their innovative long/short zero-investment factors based
on firm characteristics of size and book-to-market equity ratios sparked a
revolution in asset pricing. Loosely grounded in Merton’s (1972) ICAPM
and Ross’ APT (1976) theoretical multifactor asset pricing models, the
three-factor model became the foundation for many new factors and models
that continue to be proposed by researchers.
In this chapter we highlight some of the most popular multifactor
models in the field of asset pricing. Carhart (1997) added a momentum
factor to create a four-factor model for studying mutual fund performance.
Fama and French (2015, 2018) extended their three-factor to five- and six-
factor models. Other researchers, such as Hou et al. (2015) as well
as Stambaugh and Yuan (2017) proposed alternative four-factor models. In
a major departure from previous multifactor models, Fama and French
(2020), Lettau and Pelger (2020), and others began using machine learning
and artificial intelligence (AI) to create multifactors driven by stock return
data instead of researcher judgment. In general, asset pricing models can be
classified into those based on theoretical, discretionary, and machine
learning methods. Due to the increasing diversity of asset pricing models, a
new problem has arisen. Cochrane (2011), a factor zoo exists with hundreds
of different factors. Which factors should be used? Which combinations of
factors yield the best model? These questions are new challenges facing the
field of asset pricing.

9.1 Carhart Four-Factor Model

Momentum is an old anomaly that goes back centuries in early financial
markets around the world. The idea is simple—winner stocks (or assets)
will continue or persist to win and losing stocks will continue to lose in the
near future. Jegadeesh and Titman (1993) showed that momentum exists
for U.S. stocks in the period 1965–1989. They formed momentum
portfolios that are long stocks in the top decile of returns in the past year
and short stocks in the bottom decile of returns in the past year (see the
illustration in Fig. 9.1). This kind of relative strength trading strategy is
well known to professional investment managers. Long/short momentum
portfolios earned relatively high returns. For example, holding this portfolio
for three months yielded between 1.3 and 1.5% per month or about 16 or
17% per year! Over time these momentum profits disappear and eventually
reverse with negative earnings. Are these return patterns due to behavioral
investor responses or systematic market risk? The authors left this question
for further study.

Fig. 9.1 Momentum factor returns are long/short zero-investment portfolio returns

Carhart (1997) investigated momentum profits in mutual funds holding

diversified equity portfolios. Do mutual fund managers use momentum
persistence in stock returns to outperform other mutual funds? To answer
this question, he modified the three-factor model by incorporating a
momentum factor. His four-factor model is:
$$\begin{aligned} \begin{aligned} R_{it}-R_{ft}={}&amp;\alpha
_i+\beta _{i,m} (R_{mt}-R_{ft})+\beta _{i,S} SMB _t+\beta _{i,V}
(9.1)
HML _t \\&amp;+\beta _{i,MOM} MOM _t+e_{it}, \end{aligned}
\end{aligned}$$
where $$MOM _t$$ is the Jegadeesh and Titman momentum
factor that has low correlation with the $$SMB _t$$ and
$$HML _t$$ factors. Using mutual funds as the test assets on the left-
hand-side of Eq. (9.1) and the sample period 1963–1993, the time-series
regression model had very high $$R^2$$ values typically over 90%
and relatively low $$\alpha _i$$ values for the ith mutual fund.
Hence, the model explained differences in mutual funds’ returns. A number
of interesting findings were revealed by Carhart’s analyses.
1.
The top (bottom) decile performing mutual funds in the past year have
returns that were positively (negatively) correlated with the momentum
factor.
2.
An investor that buys last year’s top decile funds and sells the bottom
decile funds earned about 8% per year.
3.
Funds with high $$\alpha _i$$s (or mispricing errors) tended to have
above average alphas in the near future.
4.
Mutual fund management expenses, turnover (i.e., buying and selling
activity), and load fees (i.e., sales charges on investors to invest in a
fund) lowered fund returns. Due to the latter finding, no load mutual
funds with no sales charges that are passively managed (e.g., hold the S
&P 500 index) outperformed load mutual funds that are actively
managed.
5. High performing positive $$\alpha _i$$ funds (i.e., high risk-adjusted
abnormal returns) tended to have high investment expenses that offsets
their higher gross returns before expenses for the most part. More
generally, most funds actually underperformed risk-adjusted returns
after taking into account their investment expenses.
 Table 9.1 reports selected results from the Carhart study on the relation
between mutual fund performance and their operating characteristics.
Mutual funds in the low return decile in the sample period had relatively
high expense ratios and turnover compared to most other funds. Other
characteristics, such as age and size of funds, did not explain the difference
between high and low return mutual funds. Based on these and other
findings, Carhart (1997, p. 81) made the following investment
recommendations:

(1)
Avoid funds with persistently poor performance;
(2)
funds with high returns last year have higher-than-average
expected returns next year, but not in years thereafter; and
(3)
the investment costs of expense ratios, transactions costs, and
load fees all have a direct, negative impact on mutual fund
performance.

An important takeaway for investors is to buy passive mutual funds rather

than actively managed funds to earn higher risk-adjusted long-run returns.
The Carhart study demonstrates that asset pricing models can be used as
practical investment tools. Due to this study, many later models have
included a momentum factor to better explain stock returns.
Table 9.1 The relation between mutual fund performance and their operating characteristics in
the Carhart (1997) study

Age of firm Net assets Expense ratio Turnover Maximum load

High return decile 14.1 128.7 1.22 92.9 4.18
Decile 3 17.3 194.3 1.10 76.3 4.72
Decile 5 18.3 185.9 1.09 68.4 4.71
Decile 7 18.3 169.7 1.14 62.2 4.50
Low return decile 13.6 77.1 1.92 81.4 4.38

This table reports selected results from Carhart (1997). Based on the
returns of U.S. equity mutual funds, mispricing error is regressed on
different operating characteristics of mutual funds. The dependent
variable in these time-series regressions is monthly mispricing error or
$$\alpha$$ on an out-of-sample basis. More specifically, the four-factor
model is estimated for a mutual fund using three previous years of monthly
returns. Alpha is estimated in the next month using the factor loadings of
the model from the previous three years multiplied by the four factors in the
next month. This out-of-sample monthly alpha is regressed on mutual fund
characteristics. The average results for over 1,800 mutual funds are shown
in the table by mutual fund returns sorted into deciles from high to low
returns. The results shown below are based on Carhart (1997, Table IV, p.
66)

Of course, a criticism of momentum, like the size and value factors, is

the lack of theory. Why do relative strength measures explain stock returns?
They are so simple and widely used in investment practice that investors
should have arbitraged away any momentum profits in an efficient market
(e.g., buy underpriced winners and sell overpriced losers). Since momentum
consistently is significant in explaining stock returns, there must be a risk
story. What is the underlying equilibrium process in a mean-variance
investment opportunity world that accounts for its persistence as a
significant asset pricing factor? In this regard, it is worth noting that
momentum is considered by many researchers to be the foremost
anomaly or puzzle in asset pricing.
Another problem is that momentum crashes sometimes occur. Research
by Barroso and Santa-Clara (2015), Daniel and Moskowitz (2016), and
others has shown that the momentum factor becomes negative at times,
especially is down market states related to an economic recession when
market volatility is high. For this reason they scale (or divide) investments
in momentum portfolios using measures of market volatility to decrease the
magnitude of momentum portfolio return crashes. Perhaps researchers
should scale the momentum factor in asset pricing models to take into
account market conditions. More generally, given that factor returns can
dramatically change over time from economic boom times to slowdowns in
recessions, it seems reasonable to infer that the performance of asset pricing
models can be affected by the business cycle. In one sample period a model
may more significantly price assets compared to another period due to
fluctuating economic and related financial market conditions. In our
opinion, robust asset pricing factors and models should significantly price
assets under different economic and market conditions.

9.2 Fama and French Four- and Five-Factor

Models
Building on their three-factor model, Fama and French (2015) argued that
stock returns related to BM can be linked to the operating profit and capital
investment characteristics of firms. High BM (value) firms tend to have low
profit and investment, and vice versa for low BM firms.1 As theoretical
justification, they cited the dividend discount model 2 in which the value of
a share of stock is:
$$\begin{aligned} P_{it}=\sum _{\tau=1}^{\infty }
(9.2)
\frac{E(D_{t+\tau})}{(1+R)^\tau}, \end{aligned}$$
where expected future dividends per share (over an infinite time horizon)
are discounted at the required rate of return R by investors. Assuming two
stocks have the same expected dividends, stocks with higher prices have
lower discount rates over time. As risk increases, the discount rate R
increases and thereby lowers the stock price (i.e., risk-averse investors will
buy a stock at a lower price as compensation for higher risk, all else the
same). They reasoned that the expected dividends in the numerator are a
function of firm cash flows, equity earnings available to shareholders, and
changes in the book value of equity.
Based on these basic concepts, they specified the following five-factor
model:
$$\begin{aligned} \begin{aligned} R_{it}-R_{ft}={}&amp;\alpha
_i+\beta _{i,m} (R_{mt}-R_{ft})+\beta _{i,S} SMB _t+\beta _{i,V}
(9.3)
HML _t\\&amp;+ \beta _{i,R} RMW _t+\beta _{i,C} CMA _t+e_{it},
\end{aligned} \end{aligned}$$
where $$RMW _t$$ is of the robust profit minus weak profit factor,
$$CMA _t$$ is the conservative minus aggressive capital investment
factor, and other notation is as before. These new multifactors are formed
by using four sorts: two size groups, two BM groups, two operating profit
groups, and two capital investment groups. The intersection of these four
sorts gives 16 value-weighted portfolios. For example, the size factor is
long the eight small size portfolios and short the eight big portfolios.
Similarly, value, profit, and investment factors are constructed. Fama and
French were careful to point out that, in the context of the ICAPM, the four
multifactors measure sensitivities to unknown state variables. Despite not
knowing the underlying state variables, the factors capture their effects on
expected stock returns.
In the sample period 1963–2013, the average monthly risk premiums for
their five factors were as follows: $$R_m-R_f = 0.50\%$$,
$$SMB = 0.30\%$$, $$HML = 0.30\%$$,
$$RMW = 0.25\%$$, and $$CMA = 0.14\%$$. The correlations
between the multifactors were relatively low, which means that each of their
factors accounts for different information in the market. They compared the
five-factor model to different combinations of factors in three- and four-
factor models (or seven models altogether). In addition to 25 size and BM
sorted portfolios as test assets, they formed 25 size and operating profit, 25
size and capital investment (denoted Inv), 32 size and BM and operating
profit, 32 size and BM and capital investment, and 32 size and operating
profit and capital investment portfolios. They did not use industry
portfolios in their tests. Next, Gibbons, Ross, and Shanken (GRS) (1989)
tests of the joint hypothesis that all $$\alpha _i$$s equal zero for each
set of test asset portfolios were conducted. They found that:
1.
average intercepts for the five-factor model were lower than the three-
factor model; and
2.
the five-factor model performance was similar to a four-factor
model that drops the $$HML$$ factor.
Given these findings, they dropped the $$HML$$ value factor and
proposed the following more parsimonious four-factor model:
$$\begin{aligned} \begin{aligned} R_{it}-R_f={}&amp;\alpha
_i+\beta _{i,m} (R_{m}-R_{ft}) + \beta _{i,S} SMB _t+ \beta _{i,R}
(9.4)
RMW _t \\&amp;+ \beta _{i,C} CMA _t+e_{it}. \end{aligned}
\end{aligned}$$
They did not add a momentum factor due to correlations between other
multifactors. Also, more or less than four factors appeared to diminish the
performance of the multifactor model. Lastly, Fama and French (1995, p.
21) mentioned that their model had difficulty pricing some portfolios:

The portfolios of small and big stocks in the lowest B/M quartile
and highest Inv (growth stocks that invest a lot) produce intercepts
more than 3.5 standard errors from zero but of opposite sign –
(-0.20% per month, $$t = -4.18$$) for small stocks and positive
(0.37%, $$t = 5.39$$) for big stocks.

Hence, these portfolios are not priced by their four-factor model, which
suggests that a missing factor is needed to price these test assets.

9.3 Hou, Xue, and Zhang Four-Factor Model

Hou et al. (2015) developed a model that is very close to the Fama and
French four-factor model. Here we see an example of research similarities
between independent researchers. They contended that the three-factor
model was not able to account for a number of asset pricing anomalies. To
price stock portfolios associated with these anomalies, the following four-
factor model was proposed:
$$\begin{aligned} \begin{aligned} R_{it}-R_{ft}={}&amp;\alpha
_i+\beta _{i,m} (R_{mt}-R_{ft}) + \beta _{i,S} ME _t + \beta _{i,R}
(9.5)
ROE _t \\&amp;+ \beta _{i,C} IA _t + e_{it}, \end{aligned}
\end{aligned}$$
where $$ME$$ is equity market capitalization, $$ROE$$ is return on
equity, $$IA$$ is the investment to assets ratio of the firm, and other
notation is as before. Their conceptual framework is the production
(Cochrane 1991). An investment-based asset pricing model is built on the
q-theory of investment, which predicts that firms with high investment/asset
ratios (profit/assets ratios) should earn higher stock returns than those with
low investment/asset (profit/asset) ratios. As Fama and French (2015)
argued, it seems natural that capital investment and profitability impact firm
cash flows and dividends and, in turn, stock returns.
 Using U.S. stock return data from the period 1972 to 2012, their so-
called q-factor model explained many different anomalies in stock returns
(i.e., their test asset portfolios), including momentum, value versus growth,
capital investment, profitability, intangibles (e.g., corporate governance),
and trading frictions (e.g., illiquidity, idiosyncratic volatility, short-term
reversal, etc.). Of almost 80 anomalies, about one-half had insignificant
average returns, which means that these anomalies were not present in their
data. Among 35 anomalies that had significant average returns, the q-factor
model had lower average intercept ( $$\alpha _i$$) terms based on
Eq. (9.5) than the Fama and French three-factor model as well as
Carhart four-factor model—namely, 0.20%, 0.55%, and 0.55% for these
models, respectively. Only 5-out-of-35 anomalies had significant alphas (at
the five percent level) compared to 19 for the Carhart model and 27 for the
Fama and French model. They inferred that their model did a better job of
explaining anomalies than these other popular models. They did not test the
Fama and French five-factor model that appeared in the literature at the
same time in 2015.3 Hou et al. (2015, p. 685) concluded:

Many seemingly unrelated anomalies turn out to be different

manifestations of the investment and profitability effects. These
empirical findings highlight the importance of understanding the
driving forces behind the q-factors and their broad empirical
power in the cross-section.

9.4 Stambaugh and Yuan Four-Factor Model

Multifactor models continued to evolve. Stambaugh and Yuan (2017)
developed a new four-factor model as follows:
$$\begin{aligned} \begin{aligned} R_{it}-R_{ft}={}&amp;\alpha
_i+\beta _{i,m} (R_{mt}-R_{ft}) + \beta _{i,S} SMB _t + \beta
(9.6)
_{i,MT} MGMT _t \\&amp;+ \beta _{i,P} PERF _t+e_{it},
\end{aligned} \end{aligned}$$
where $$MGMT$$ is a management factor, $$PERF$$ is a
performance factor, and other notation is as before. The latter
$$MGMT$$ and $$PERF$$ factors were constructed from stocks’
rankings with respect to a number of asset pricing anomalies. They grouped
11 well-known anomalies4 into two clusters. Next, these clusters were used
to define two long/short zero-investment portfolios (viz., MGMT and
PERF), which they called mispricing factors. Unlike other models, their
size factor is not based on a two-way sort on size and BM. Departing from
previous researchers, they drop stocks contained in their two mispricing
factors’ portfolios and create a small-minus-big size factor. These stocks are
less mispriced than other stocks. Further departing from other models, they
do not include a value factor (HML). They found that their mispricing
factors MGMT and PERF priced BM portfolios such that the value factor
was redundant in their model.
Table 9.2 Intercept or $$\alpha$$ estimates for selected anomaly portfolios from January 1967 to
December 2013 in the Stambaugh and Yuan (2017) study

Anomaly return FF 3-factor FF 5-factor HXZ 4-factor q SY 4-factor

Net stock issues 0.56*** 0.66*** 0.32*** 0.37*** 0.06
Accruals 0.43*** 0.51*** 0.56*** 0.65*** 0.31**
Asset growth 0.52*** 0.32*** 0.06 0.08 −0.22**
Distress 0.44 1.21*** 0.62** 0.20 −0.16
Momentum 1.26*** 1.59*** 1.35*** 0.48 0.12
Gross profitability 0.28* 0.69*** 0.35*** 0.39** 0.27
Book-to-market 0.43** −0.20** −0.14 −0.03 −0.17

This table reports selected results from Stambaugh and Yuan (2017). Based
on U.S. stock returns, monthly $$\alpha$$ estimates (in percent) are
shown. These mispricing errors are estimated for the following models:
Fama and French three- and five-factor models, Hou, Xue, and Zhang four-
factor q model, and Stambaug and Yuan four-factor model. The sample
period is January 1967–December 2013. Anomaly returns are long-short
portfolio returns using high and low decile returns. The results shown
below are based on Stambaugh and Yuan (2017, Table 4, p. 1287)
Note Asterisks indicate that the $$\alpha$$ estimate is statistically
different from zero at the
* = 10%, ** = 5%, and *** = 1% levels

Using the sample period 1967–2013 and U.S. stock returns, they tested
alternative models using the anomalies in Hou et al. (2015). Table 9.2
shows some selected results from their study. In the left-hand column of
numbers, the table gives the average anomaly return computed as long
minus short portfolio returns using high and low decile returns for each
anomaly. They compared the multifactor model in Eq. (9.6) with the Fama
and French three- and four-factor models and Hou, Xue, and Zhang four-
factor q-factor model. Notice that estimated alphas ( $$\alpha _i$$s)
tend to be lower in magnitude and significance (indicated by asterisks) for
the Hou, Xue, and Zhang q-factor model as well as Stambaugh and
Yuan four-factor models compared to the Fama and French three- and four-
factor models. Also, comparing the four-factor models in the last two
columns, their performance is somewhat similar, but the Stambaugh and
Yuan model outperforms the q-factor model in terms of smaller and less
significant mispricing error.
The authors reported further evidence that their four-factor model did a
better job of pricing factors in the other models. That is, they used the
factors in other models as anomaly portfolios or test assets. In view of these
and other tests, they concluded that, instead of constructing each factor
from a single anomaly, multifactor models can be improved by developing
factors from multiple anomalies. In this way each factor has a broader
ability to price a wider array of stocks manifesting different anomalies. This
innovative multifactor approach has the advantage of reducing the number
of factors in asset pricing models. With so many factors being proposed by
researchers that are associated with different anomalies, Stambaugh and
Yuan (2017, p. 1307) observed:

In this sense, while parsimonious factor models are appealing and

useful, there are limits to their abilities to explain expected returns.
Nevertheless, among such models, those with mispricing factors
appear to have greater ability than prominent alternatives.

As additional proof, they constructed a single mispricing factor from the 11

anomalies and found that a three-factor model with market, size, and the
mispricing factor outperformed the Fama and French five-factor model. In
other words, the value, profit, and capital investment factors can be
collapsed into one mispricing factor using their methods. This multiple
mispricing approach to constructing factors therefore can help to reduce the
number of factors to develop more parsimonious asset pricing models.
9.5 Fama and French Six-Factor Model
Continuing to build on their earlier empirical work, Fama and French
(2018) added the momentum factor (MOM) as in Carhart (1997) to their
five-factor model to form a six-factor model5:
$$\begin{aligned} \begin{aligned} R_{it}-R_{ft}={}&amp;\alpha
_i + \beta _{i,m} (R_{mt}-R_{ft}) + \beta _{i,S} SMB _t + \beta
(9.7)
_{i,V} HML _t \\&amp;+ \beta _{i,R} RMW _t + \beta _{i,C} CMA
_t+\beta _{i,M} MOM _t+e_{it}, \end{aligned} \end{aligned}$$
where notation is as before. Moreover, they tested other value, profitability,
and capital investment factors by further sorting each factor into a small and
big factor (e.g., small stocks with high BM returns minus small stocks with
low BM returns = $$HML _S$$ and big stocks with high BM returns
minus big stocks with low BM returns = $$HML _B$$). Using sorts on
size with the other factors (not just BM as in the past), a number of different
size factors were created also. Altogether, excluding the market factor, they
considered 48 multifactors!
The empirical analyses sought to determine which configuration of
factors provided the best model. To do this, various tests of the intercepts (
$$\alpha _i$$s) were conducted using U.S. stock returns from 1963 to
2016. The winner model among different models was the one with the
smallest squared Sharpe ratio for the intercepts.6 Different versions of their
three-, five-, and six-factor models were tested using various combinations
of multifactors. They found that the winner was a six-factor model
comprised of market and size factors plus small stock factors for the value,
profitability, capital investment, and momentum factors. Other tests rejected
the CAPM as well as three- and five-factor models.7 With so many factors
and contender models, it seems reasonable to ask: How should researchers
choose factors and models? To avoid potential data snooping (or dredging),
Fama and French (2018, p. 247) gave the following cautionary advice:

We suggest that model comparisons in any paper should be limited

by theory, even an umbrella theory like the dividend discount
model, and by evidence on model robustness out-of-sample
(different time periods and markets).
As mentioned before, intercept tests are typically in-sample, not out-of-
sample tests in which the model is estimated in a sample period and then
tests are conducted on returns in the subsequent period.8 For out-of-sample
tests, Fama and MacBeth (1973) two-step regression tests are needed. We
will return to this discussion shortly.
9.6 Fama and French Cross-Section Factors
Recently, Fama and French (2020) proposed modified versions of their
models in which the multifactors are replaced by long/short mimicking
portfolios estimated from a cross-sectional regression. As observed by
Ferson (2019, p. 223) and others, estimated coefficients (or
$$\lambda _k$$s) in a Fama and MacBeth cross-sectional regression are
long/short portfolios (e.g., long high BM stock returns and short low BM
stock returns for the market price of risk associated with value factor
loadings). Given this fact, Fama and French used these mimicking
portfolios as the factors in their five-factor model in Eq. (9.3). Let’s take a
closer look at their methods.
In one version of their new model, using a cross-sectional
regression with tests asset portfolios, they regress portfolios’ returns in
month t (e.g., 25 size and BM portfolios) on lagged firm characteristic
values at time $$t-1$$ as follows:
$$\begin{aligned} \begin{aligned} R_{it}={}&amp;R_{zt}+R_{
MC t} MC _{it-1} + R_{ BM t} BM _{it-1} + R_{ OP t} OP _{it-1}
(9.8)
\\&amp;+ R_{ INV t} INV _{it-1}+e_{it}, \end{aligned}
\end{aligned}$$
where $$MC _{it-1}$$, $$BM _{it-1}$$, $$OP _{it-1}$$,
and $$INV _{it-1}$$ are observed values of size, the book-to-market
ratio, operating profitability, and the rate of growth of assets. The intercept
$$R_{zt}$$ was named the level return not explained by these firm
characteristic variables. In this equation, the corresponding
$$\lambda _k$$ coefficients (associated with the k factor loadings)
are denoted $$R_{ MC t}$$, $$R_{ BM t}$$,
$$R_{ OP t}$$, and $$R_{ INV t}$$, respectively. These
$$\lambda _k$$s are the mimicking portfolio returns or so-called
cross-section factors. They then created a new five-factor model by
substituting these cross-section factors into their original time-series five-
factor model in Eq. (9.3) as follows:
(9.9)
 $$\begin{aligned} \begin{aligned} R_{it}-R_{ft}={}&amp;\alpha
_i+\beta _{i,m} (R_{mt}-R_{ft})+\beta _{i,S} R_{ MC t}+\beta
_{i,V}R_{ BM t} \\&amp;+\beta _{i,R}R_{ OP t}+\beta _{i,C}R_{
INV t}+e_{it}, \end{aligned} \end{aligned}$$
where their long/short factors $$SMB$$, $$HML$$,
$$RMW$$, and $$CMA$$ are replaced by the cross-section
factors.
Previous studies by Back et al. (2013, 2015) showed that this cross-
sectional regression approach is a way to let the data itself (i.e., accounting
variables for firms) determine multifactors rather than constructing them
using 2 $$\,\times \,$$3 size and BM sorts (see Fig. 8.1) that are ad
hoc or subjective in terms of using researcher judgment. We can consider
this new five-factor model a machine learning competitor of their original
five-factor model that was based on researcher judgment.
In another twist, they proposed a mathematical four-factor model as
follows:
$$\begin{aligned} \begin{aligned} R_{it}-R_{zt} ={}&amp;MC
_{it-1}R_{ MC t} + BM _{it-1}R_{ BM t} + OP _{it-1}R_{ OP t}
(9.10)
\\&amp;+ INV _{it-1}R_{ INV t}+e_{it}, \end{aligned}
\end{aligned}$$
where there is no intercept mispricing term, and the “coefficients” are the
accounting variables lagged one period. You can see that there are no
regression coefficients in this model. It is simply a mathematical
model with time-varying loadings (accounting variables) in place of
constant beta coefficients. The cross-section factors are the multifactors.
There is no market factor which would require a regression coefficient to be
estimated if it were in the model. The average of the error terms
$$e_{it}$$ over t time periods becomes the mispricing alpha in this
model.9
The advantage of the mathematical model, in the spirit of conditional
models discussed in Chapter 6, is that the risk loadings are allowed to vary
over time. Recall earlier that, since we know that beta risks vary over time,
it was hoped that conditional models could strengthen the CAPM and other
models. However, research showed that the improvement was not sufficient
to save the CAPM or materially improve the results of other models.
Perhaps this new conditional model using time-varying accounting
variables could revive the conditional approach.
Fama and French specified a second conditional model with time-
varying risk loadings using their five-factor model:
$$\begin{aligned} R_{it}-R_{ft}={}&amp;\alpha _i + \beta _{i,m}
(R_{mt}-R_{ft}) + \beta _{i,S} SMB _t + \beta _{i,V} HML _t +
\beta _{i,R} RMW _t \\&amp;+ \beta _{i,C} CMA _t\nonumber +
(9.11)
\beta _{i,S2} MC _{it-1} SMB _t + \beta _{i,V2} BM _{it-1} HML
_t \\&amp;+ \beta _{i,R2} OP _{it-1}RMW_t \nonumber + \beta
_{i,C2} INV _{it-1} CMA _t + e_{it}, \end{aligned}$$
where the interaction terms (e.g., $$MC _{it-1}\times SMB _t$$)
allow for conditional time variation in the regression coefficients. The
accounting variables $$MC$$, $$BM$$, $$OP$$, and
$$INV$$ are the instruments in this regression model. They did not
instrument the intercept, which therefore is constant over time. Will this
model perform as well as the conditional mathematical model?
U.S. stock returns from 1963 to 2018 were employed in their analyses.
They used joint tests that the intercepts ( $$\alpha _i$$s) equal zero.10
The main findings of their study were:
1.
conditional models (9.10) and (9.11) outperformed the other models
with constant risk loadings over time;
2.
conditional mathematical model (9.10) outperformed conditional model
(9.11) with instruments;
3.
cross-section factor model (9.9) (with constant coefficient loadings)
outperformed conditional model (9.11) with instruments to allow time-
varying risk loadings; and
4.
cross-section factor model (9.9) (with constant coefficient loadings)
outperformed their original five- and six-factor models but only by a
small or marginal amount.
From these results, Fama and French (2020, p. 1915) concluded:
 Simply substituting CS (cross-section) factors for TS (time-series)
factors in regression models ... doesn’t improve much on the
descriptions of average returns.

In other words, the cross-section factors did not improve their original time-
series models enough to warrant their use. However, using cross-section
factors in combination with firm characteristics (or accounting variables) in
the conditional mathematical model improved upon time-series models
using researcher definitions of factors based on firm characteristics as in
their original models.
A drawback of the conditional mathematical model is that no out-of-
sample Fama and MacBeth tests are possible. The intercept tests performed
in their study are entirely in-sample tests. Simin (2008) has argued that step
ahead (e.g., one-month-ahead that are out-of-sample) Fama and MacBeth
tests avoid a number of model evaluation problems, such as data snooping.
Also, Ferson et al. (2013) have recommended that the practical value of
asset pricing models should be evaluated based on out-of-sample tests.
An important point here is that model tests should be investable. For
example, an investor picks stocks based on the estimation of their risk and
then subsequently tracks their return performance in the future. Fama–
MacBeth tests first estimate model parameters from available historical
data. In the second step, returns in the next (out-of-sample) month are
regressed on previously estimated parameters (beta loadings). There is no
possibility of cheating in this setup. The model that performs best in the
second step dominates the other models. Note that some researchers use an
in-sample Fama–MacBeth approach. Model parameters are estimated with
monthly returns over many years in a sample period. Then returns in each
month within the sample period are regressed on the model parameters in
the second step. But this method is not investable. Investors could not rely
on this kind of analysis to guide them concerning the relation between past
risk and future returns. They want models that give estimates of risk
parameters that are related to returns in the future rather than the past. The
Fama and French conditional mathematical model (using time-varying firm
characteristics as risk loadings and cross-section factors) proved itself to be
a dominant model using in-sample tests. However, out-of-sample tests are
needed to assess its practical use as an investable strategy.
9.7 Lettau and Pelger Latent Five-Factor Model
In light of the growing number of possible factors in asset pricing models,
Lettau and Pelger (2020) proposed a method for finding the most important
factors. Their method identifies latent asset pricing factors using Principal
Component Analysis (PCA). PCA is a statistical method that can reduce the
information contained in sample observations (e.g., stock portfolios’
returns) into a select few components that summarize all of the sample
information.
Each factor in an asset pricing model can be considered a risk
dimension. The CAPM is a single risk dimension (market factor) model.
The three-factor Fama and French model has three-dimensional risk. If say
50 different stocks are input into a PCA (i.e., we start with 50 dimensions),
the PCA seeks to reduce the dimensions down to a smaller set that explains
the data. In this way PCA is very similar to factor analysis covered in
Chapter 7’s discussion of empirical tests of Ross’ (1976) APT. It is
reasonable to believe that some of the stocks’ returns are correlated with
one another. For example, 10 stocks’ returns could be highly correlated and
therefore are combined into one dimension. The first dimension in PCA has
higher explanatory power in terms of stock returns than other dimensions.
In their study, they found that stock returns could be explained by a five-
dimensional space. The sixth dimension and higher dimensions were
dropped because they did not significantly help to explain the data
observations (or stock returns).
Since the dimensions are not readily observable, they are referred to as
latent or hidden. Each component is constructed by PCA to be orthogonal
(or uncorrelated) with other components. For this reason, PCA has some
connection to the intertemporal CAPM (ICAPM) of Merton (1972) and
arbitrage pricing theory (APT) of Ross (1976) which assume orthogonal
asset pricing factors.
PCA is a machine learning technique that has been applied in many
fields of study, including finance and economics. Rather than use researcher
judgment to form factors, computer algorithms within the PCA software let
the data select the factors based on correlation patterns in stock returns. In
Fama and French’s (2020) study of cross-section factors, they let the data
compute these factors from cross-sectional regressions. PCA is another
example of machine learning in factor construction. According to Lettau
and Pelger, PCA analyses yielded five systematic factors.
In the sample period 1963–2017, the authors gathered U.S. stock returns
for 370 portfolios sorted on size, BM, accruals, investment, profitability,
momentum, volatility measures, and other characteristics of firms. In-
sample intercept ( $$\alpha _i$$) tests of mispricing errors were
performed on different models. Some out-of-sample intercept tests were
conducted also (i.e., factor loadings were estimated and then average
pricing errors were computed in the step ahead period).
Table 9.3 Mispricing error of different models in the Lettau and Pelger (2020) study

In-sample Out-of-sample In-sample Out-of-sample

Model 370 portfolios 370 portfolios 270 stocks 270 stocks
LP-PCA 0.16 0.15 0.27 0.43
PCA 0.22 0.19 0.28 0.41
FF three-factor 0.31 0.25 0.43 0.43
FF five-factor 0.26 0.19 0.39 0.42

This table reports selected results of mispricing error tests in Lettau and
Pelger (2020). Average mispricing error is measured by the root mean
square of alpha pricing error computed as follows:
$$RMS _\alpha =(\hat{\alpha }^{T}\hat{\alpha }/N)^{1/2}$$, where
$$\alpha$$ is mispricing error for T months, and N is the number of
portfolios or stocks. There are four models: the Lettau-Pelgar five-factor
PCA model, five-factor PCA model based on previous methods, and Fama
and French three- and five-factor models. In-sample and out-of-sample
$$RMA_\alpha$$ estimates are shown for 370 stock portfolios and 270
individual stocks. The sample period is from November 1963 to December
2017 ( $$T=650$$). The results shown below are based on Lettau and
Pelger (2020, Table 2, p. 2297)

Table 9.3 summarizes selected results in their paper. They compared

their PCA model (denoted LP-PCA) to standard PCA methods as well as
the Fama and French three- and five-factor models. The root mean square
of alpha ( $$\alpha$$) estimates, or $$RMS _\alpha$$, was used
to measure mispricing error in the sample period comprised of 650 months.
Root mean square error (RMSE) is a common measure used in statistics to
compare the forecasting errors of different models. Using 370 stock
portfolios, mispricing errors were smaller for the Lettau and Pelger PCA
model (LP-PCA) than the other models both in- and out-of-sample (e.g.,
0.16 and 0.15, respectively, compared to 0.26 and 0.19 for the Fama and
French five-factor model). Using 270 individual stocks, LP-PCA
outperformed the other models in terms of in-sample mispricing error but
had similar out-of-sample mispricing error. Summarizing their results, the
evidence indicated that the proposed latent five-factor PCA model had
smaller pricing errors compared to other models.
Interestingly, the five PCA latent factors identified by Lettau and Pelger
were correlated with well-known multifactors, including market, value, and
momentum factors. For example, the first latent factor was highly
correlated with the value-weighted CRSP market index. This correlation
evidence suggests that the latent factors are related to the same systematic
risk factors as popular multifactors. Lettau and Pelger (2020, p. 2274)
inferred that:

Our factors imply that a significant amount of characteristic

information is redundant.

In other words, machines can do a good job of capturing underlying risk

factors in stock returns. As such, researcher judgment is not needed about
which firm characteristics should be used in multifactor models. For
interested readers, another recent study using PCA is Kozak et al. (2020),
who employed portfolios based on firm characteristics to choose a few
principal component factors that achieved good out-of-sample performance.
While PCA methods yield significant factors, the sources of economic risk
underlying machine-made factors can be difficult to discern.

9.8 Rise of the Machines

Cross-section factor models in Fama and French (2020) and latent factor
models in Lettau and Pelger (2020) and others break new ground in asset
pricing. Both types of models employ a machine learning approach to
develop empirical asset pricing models. Less researcher judgment and
subjective factor construction is used in these mechanical types of models.
Large amounts of data are mined by fairly sophisticated computer
algorithms and statistical programs in an effort to detect useful patterns in
asset pricing. The computer essentially constructs factors by analyzing large
amounts of data, thereby moving asset pricing closer to artificial
intelligence (AI). Previous multifactor models utilized what we can refer to
as discretionary methods that depend on human judgment rather than
machine learning methods. Figure 9.2 summarizes the multifactor models
(i.e., discretionary methods). Figure 9.3 reviews the machine learning
models (i.e., AI methods).

Fig. 9.2 Multifactor models extensions of the original Fama and French three-factor model using
discretionary methods

Fig. 9.3 Machine learning methods based on artificial intelligence (AI) models rather than human
judgment

A missing link in both discretionary and machine learning methods is

theory. The CAPM, its companion models, and the APT relied heavily on
efficient markets, equilibrium pricing, and portfolio diversification. These
traditional theoretical methods have been supplanted by contemporary
discretionary and machine learning methods due to their better data fitting
ability with respect to stock returns. In the next chapter, we will review a
new model dubbed the ZCAPM that falls within the wheelhouse of
theoretical methods. As we will see, the ZCAPM appears to be a serious
contender in the pantheon of asset pricing models due to its strong
empirical ability to explain the cross-section of average stock returns. Also,
it provides a missing link that grafts discretionary and machine learning
methods onto the traditional theoretical branch of asset pricing methods.

9.9 Summary
After extensive tests using U.S. stock returns over many years, Fama and
French (1996) declared the CAPM dead. Little or no relation between U.S.
stock returns and beta risk led them to propose new multifactor models. In
their papers, Fama and French (1992, 1993, 1995, 1996) introduced a three-
factor model, which proved to boost the empirical fit to stock returns of the
CAPM’s market factor by using the firm characteristics of size and value
(i.e., book-to-market equity ratios denoted BM). A major innovation was
the construction of long/short zero-investment factors—for example, long
small stocks and short big stocks as well as long high BM stocks and short
low BM stocks. These so-called multifactors triggered a revolution in asset
pricing.
The multifactor movement resulted in many new long/short factors
proposed by researchers. Carhart (1997) added a momentum factor to the
three-factor model to make a four-factor model. The momentum factor is
long stocks with high past returns in the last year and short stocks with low
past returns. This new factor was helpful in explaining the returns on
mutual funds, which are portfolios of securities. Practical implications for
investment were gained, such as to buy funds with higher-than-average
performance in the recent past. Also, buy passively managed index funds
rather than actively managed funds with higher operating expenses. Fama
and French (2015) augmented their three-factor model with two more
factors—namely, profit and capital investment factors—to make a five-
factor model. These additional long/short factors further boosted the
empirical fit and suggested that the value factor could be dropped.
Other researchers began to enter the multifactor competition. Hou et al.
(2015) advanced a four-factor model with the market factor plus new size,
profit, and capital investment factors. They linked their model to production
CAPM models by means of the q-theory of investment. According to this
theory, firms with high investments and profits should tend to earn higher
stock returns, and vice versa for those with low investments and profits.
They found that their model helped to explain 80 anomalies, most of which
were based on different firm characteristics. In this respect, their model
outperformed the Fama and French three-factor and Carhart four-factor
models. They inferred that many of these anomalies in stock returns were
effectively explained by investment and profit factors. Another study by
Stambaugh and Yuan (2017) proposed an alternative four-factor
model containing market and size factors plus management and
performance mispricing factors. To construct their two mispricing factors,
they formed two clusters of a number of anomalies and then computed
long/short factors from them. Their model was able to do a good job of
explaining 80 different anomalies and exceeded the performance of the
Fama and French five-factor model and Hou, Xue, and Zhang four-factor
model. Hence, an effective way to construct a parsimonious set of factors is
from the anomalies themselves.
Extending their previous work, Fama and French (2018) put forward a
six-factor model that added momentum to their five-factor model. In this
model they defined long/short factors in different ways to test a variety of
size, value, profit, capital investment, and momentum factors. The winner
model contained market and size factors plus small stock factors for size,
value, capital investment, and momentum. In a related paper, Fama and
French (2020) proposed the use of cross-section factors that are estimated
from cross-sectional market prices of risk. That is, in the second step of
the Fama and MacBeth (1973) procedure to test a model, the cross-
sectional regression produces $$\lambda _k$$ estimates associated
with the factor loadings for k factors. These $$\lambda _k$$s are
mimicking portfolio returns dubbed cross-section factors that were
substituted into their five- or six-factor models in place of the original
factors. Another innovation was to specify a conditional mathematical
model with time-varying firm characteristics as the loadings on the cross-
section factors in a mathematical conditional model. This mathematical
model outperformed the regression models tested in their study. In effect,
cross-section factors use machine learning instead of researcher judgment to
design factors.
In another recent study using machine learning, Lettau and Pelger
(2020) applied Principal Components Analysis (PCA) to search for
correlation patterns in stock returns that represent so-called latent factors.
They found that five latent factors were suggested by five-dimensional risk
patterns in stock returns. The first factor was highly correlated with the
CRSP market index, and other factors were related to value and momentum
factors. Tests with stock portfolios associated with well-known asset pricing
anomalies, including size, BM, investments, profitability, momentum, and
others, showed that their five-factor model had small mispricing errors (or
$$\alpha _i$$s) and therefore explained stock returns.
In sum, theoretical methods led to the creation of the CAPM, its
companion models, and the APT. Over time, these models were supplanted
by better fitting models based on discretionary methods that emphasized
researcher judgment, including the three-, four-, five-, and six-factor
models. Researchers constructed various long/short zero-investment
portfolios as multifactors that were added to the CAPM market factor. More
recently, machine learning methods have been proposed that produce
factors based on artificial intelligence (AI). Cross-section and latent
factors employ statistical and computer algorithms to form factors. In this
man versus machine battle, who will win? Will some hybrid approach
emerge in the future? As we will see in the next chapter, an unlikely
outcome may be a return to theoretical methods with a new model dubbed
the ZCAPM.

Questions
1.
What is momentum in asset returns? Can it be used to construct a
long/short factor? How did Carhart (1997) modify the three-factor
model to take into account momentum?
2.
In Carhart’s (1997) empirical tests, did momentum affect stock returns
of mutual funds? What practical advice to investors in mutual funds
did his research reveal?
3.
Fama and French (2015) built upon their three-factor model to
develop a five-factor model. What motivated them to add two more
factors to their model? Write their five-factor model.
4.
Fama and French (2015) tested their five-factor model in the sample
period 1963–2013. What test asset portfolios did they use? Was the
five-factor model supported?
5.
Hou et al. (2015) developed a model that is very close to the Fama
and French four-factor model. Write the equation for their model.
What is the theoretical foundation for their model? What did they use
as the test assets to empirically test their model? What did they find?
6.
Stambaugh and Yuan (2017) developed a four-factor model. They
added two new factors to the market and size factors. How did they
construct these new factors? In empirical tests of over 80 anomalies,
what did they find? They also created a three-factor model. How did
this model perform?
7. Write the Fama and French (2018) six-factor model. They expanded
this model to include different variations of these factors for a total of
 this model to include different variations of these factors for a total of
48 factors. Which model worked the best?
8.
More recently, Fama and French (2020) proposed a modified version
of their five-factor model in which the multifactors are replaced by
long/short mimicking portfolios estimated from a cross-sectional
regression. What is the two-step process they implemented to do this?

9.
Fama and French (2020) proposed two additional models that allow
for time-varying risk parameters. Write the equations for these two
conditional models. In empirical tests, which model was the best one?
In these tests, how did they test one of the models that had no
intercept (mispricing $$\alpha$$) term?
10.
Normally, joint tests that the intercepts ( $$\alpha _i$$s) equal zero
are in-sample tests. What is a criticism of this testing approach?
11.
Lettau and Pelger (2020) proposed the use of Principal Component
Analysis (PCA) to identify latent asset pricing factors. How does PCA
find factors in stock returns? How many latent factors did they find in
empirical tests? Are PCA latent factors related to traditional factors
used in asset pricing models?
12.
Distinguish between theoretical methods, discretionary methods, and
machine learning methods of developing asset pricing models. What
are examples of each of these methods?

Problems
1.
Draw a diagram of returns on the Y-axis and time in months over the
past year on the X-axis. Using this diagram, illustrate how momentum
portfolios are formed.
2.
Draw a diagram comparing the machine learning methods of Fama and
French’s (2020) cross-section factors and Lettau and Pilger’s (2020)
latent factors.
References
Back, K., N. Kapadia, and B. Ostdiek. 2013. Slopes as factors: Characteristic pure plays. Working
paper, Rice University.

Back, K., N. Kapadia, and B. Ostdiek. 2015. Testing factor models on characteristic and covariance
pure plays. Working paper, Rice University.

Barillas, F., and J. Shanken. 2018. Comparing asset pricing models. Journal of Finance 73: 715–754.

Barroso, P., and P. Santa-Clara. 2015. Momentum has its moments. Journal of Financial Economics
116: 111–120.
[Crossref]

Carhart, M.M. 1997. On persistence in mutual fund performance. Journal of Finance 52: 57–82.
[Crossref]

Cochrane, J.H. 1991. Production-based asset pricing and the link between stock returns and
economic fluctuations. Journal of Finance 46: 209–237.
[Crossref]

Cochrane, J.H. 2011. Presidential address: Discount rates. Journal of Finance 56: 1047–1108.
[Crossref]

Daniel, K., and T. Moskowitz. 2016. Momentum crashes. Journal of Financial Economics 122: 221–
247.
[Crossref]

Fama, E.F., and J.D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy 81: 607–636.
[Crossref]

Fama, E.F., and K.R. French. 1992. The cross-section of expected stock returns. Journal of Finance
47: 427–465.
[Crossref]

Fama, E.F., and K.R. French. 1993. The cross-section of expected returns. Journal of Financial
Economics 33: 3–56.
[Crossref]

Fama, E.F., and K.R. French. 1995. Size and book-to-market factors in earnings and returns. Journal
of Finance 50: 131–156.
[Crossref]

Fama, E.F., and K.R. French. 1996. The CAPM is wanted, dead or alive. Journal of Finance 51:
1947–1958.
[Crossref]

Fama, E.F., and K.R. French. 2015. A five-factor asset pricing model. Journal of Financial
Economics 116: 1–22.
[Crossref]

Fama, E.F., and K.R. French. 2018. Choosing factors. Journal of Financial Economics 128: 234–252.
[Crossref]

Fama, E.F., and K.R. French. 2020. Comparing cross-section and time-series factor models. Review
of Financial Studies 33: 1892–1926.
[Crossref]

Ferson, W.E. 2019. Empirical Asset Pricing: Models and Methods. Cambridge, MA: The MIT Press.

Ferson, W.E., S.K. Nallareddy, and B. Xie. 2013. The “out-of-sample” performance of long-run risk
models. Journal of Financial Economics 107: 537–556.
[Crossref]

Gibbons, M.R., S.A. Ross, and J. Shanken. 1989. A test of the efficiency of a given portfolio.
Econometrica 57: 1121–1152.
[Crossref]

Gordon, M.J. 1962. The Investment, Financing, and Valuation of the Corporation. Homewood, IL:
R.D. Irwin.

Hou, K., C. Xue, and L. Zhang. 2015. Digesting anomalies: An investment approach. Review of
Financial Studies 28: 650–705.
[Crossref]

Jegadeesh, N., and S. Titman. 1993. Returns to buying winners and selling losers: Implications for
stock market efficiency. Journal of Finance 48: 65–91.
[Crossref]

Kozak, S., S. Nagel, and S. Santosh. 2020. Shrinking the cross section. Journal of Financial
Economics 135: 271–292.
[Crossref]

Lettau, M., and M. Pelger. 2020. Factors that fit the time series and cross-section of stock returns.
Review of Financial Studies 33: 2274–2325.
[Crossref]

Merton, R.C. 1972. An analytical derivation of the efficient portfolio frontier. Journal of Financial
and Quantitative Finance 7: 1851–1872.
[Crossref]

Novy-Marx, R. 2013. The other side of value: The gross profitability premium. Journal of Financial
Economics 108: 1–28.
[Crossref]

Ross, S.A. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13: 341–
360.
[Crossref]

Simin, T. 2008. The poor performance of asset pricing models. Journal of Financial and Quantitative
Analysis 43: 335–380.
[Crossref]

Stambaugh, R.F., and Y. Yuan. 2017. Mispricing factors. Review of Financial Studies 30: 1270–1315.
[Crossref]

Footnotes
1 Novy-Marx (2013) also found that stock returns are related to expected profits.

2 Gordon (1962) is famous for this simple valuation model.

3 In a later study by Barillas and Shanken (2018), various factors and models in previous studies
were tested. Using novel tests of the relative performance of different models, in both in-sample and
out-of-sample tests, they found that a six-factor model containing market, size, value, profit,
investment, and momentum factors outperformed other possible models.

4 The 11 anomalies were: net stock issues, equity issuance, accruals, net operating assets, asset
growth, investment to assets, financial distress, O-score of distress, momentum, gross profitability,
and return on assets.

5 As cited in footnote 3, Barillas and Shanken (2018) tested this model also.

6 Beyond the scope of the present introductory text, this squared Sharpe ratio test of intercepts was
based on work by Barillas and Shanken (2018), who developed the performance metric for
comparing competing asset pricing models.

7 Barillas and Shanken (2018) tested different factors and models and found that a six-factor model
with market, size, value, profit, investment, and momentum factors outperformed other models in
both in-sample and out-of-sample tests.

8 Fama and French (2018) conducted a different kind of out-of-sample test for models by using a
subsample of returns in the sample period for in-sample test statistics and then using the remaining
returns in the same sample period for out-of-sample test statistics.
9 In regression models with an intercept term, OLS fits the model such that the average of the
residuals, or $$E(e_{it})$$, approximately equals zero. Without an error term in model (9.10) the
average of the residual terms will not be zero and, therefore, is a good proxy for the mispricing alpha.

10 As in their 2015 study of the six-factor model, they applied maximum squared Sharpe ratio tests
of the intercepts also (see footnote 6).
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_10

10. A Special Case of the Zero-Beta CAPM: The ZCAPM

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX, USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa, Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1007/978-3-031-16784-3_10

A new CAPM dubbed the ZCAPM was recently proposed by Kolari, Liu, and Huang (KLH) (2021). In their book,
the authors derived the ZCAPM as a special case of the more general zero-beta CAPM of Black (1972).1 The
theoretical ZCAPM is comprised of only two factors: mean excess market returns and the cross-sectional standard
deviation of returns for all assets in the market more simply referred to as return dispersion. Sensitivity of asset
returns to these two factors is measured by beta risk and so-called zeta risk, respectively. Their derivation of the
ZCAPM relied heavily on the mean-variance investment parabola of Markowitz (1959), the CAPM equilibrium
framework of Treynor (1961, 1962), Sharpe (1964), Lintner (1965), and Mossin (1966), and Black’s zero-beta
CAPM with two factors.
Relative to multifactor models, empirical tests using U.S. stock returns strongly supported the special case of
the ZCAPM and, therefore, the more general zero-beta CAPM. Out-of-sample, cross-sectional Fama and MacBeth
(1973) regression tests of the factor loadings (or beta risk and zeta risk coefficients) showed that zeta risk
associated with return dispersion is more significantly priced than other factors in popular multifactor models.
Based on the sample period 1965–2018, the authors compared the ZCAPM to the CAPM, Carhart (1997) four-
factor model, Fama and French (1992, 1993, 1995, 1996, 2015, 2018), three-, five-, and six-factor models, Hou
et al. (2014) four-factor model, and Stambaugh and Yuan (2017) four-factor model. Many different test asset
portfolios (including individual stocks) were used as well as split sample tests with subperiods.
Surprisingly, the ZCAPM consistently outperformed multifactor models in terms of both higher $$R^2$$
values of goodness-of-fit and significance of the return dispersion factor. While multifactors are statistically
significant in many cases, which confirms many published studies, they are not as significant as return
dispersion in virtually all tests. Moreover, when using industry portfolios, which are exogenous to the factors in
models, the ZCAPM dominated other models by large margins in many instances. Recall that discretionary
methods incorporating multifactors in asset pricing models were initiated by Fama and French (1992, 1993) to
boost the goodness-of-fit of the market model form of the CAPM and other companion CAPM specifications. The
finding that goodness-of-fit can be improved by a CAPM-based asset pricing model is a remarkable turn of events.
In this chapter we review ZCAPM evidence in KLH’s book in addition to findings related to many of the
multifactor models covered in Chapters 8 and 9. Interestingly, KLH argue that the ZCAPM has some relation to
discretionary methods relying on long/short multifactors. Each long/short portfolio in these models (e.g., the Fama
and French three-factor model) is a rough measure of cross-sectional returncross-sectional return dispersion. For
example, the size factor is small stocks’ returns minus big stocks’ returns. This factor comes from the cross-
sectional distribution of all stock returns in the market. It represents one slice of the total return dispersion that
describes the distribution of stock returns at a point in time. Other factors, such as value, profit, capital investment,
momentum, etc., capture other slices within the total return dispersion. Together, multifactors incorporate most of
the total return dispersion. However, each multifactor factor can change over time in terms of their location within
the total return dispersion. Another interesting aspect of multifactors is that they sometimes can become negative
(e.g., big stock returns are greater than small stock returns). The ZCAPM takes into account this kind of outcome
as the empirical form of the ZCAPM allows for both positive and negative zeta risk associated with return
dispersion.
Thus, as discussed by the authors, multifactor models are related to the ZCAPM and thereby the CAPM.
Multifactor models are other forms of the ZCAPM that are more appropriately modeled using total return
dispersion rather than different slices of return dispersion. Even new models using machine learning methods that
utilize long/short portfolios as multifactors are related to the ZCAPM. It appears that the CAPM is not dead after
all but is alive and well!

10.1 Theoretical ZCAPM

To derive the ZCAPM, KLH utilized Markowitz’s mean-variance investment parabola. Figure 10.1 shows the
parabola with expected portfolio returns $$E(R_P)$$ on the Y-axis and the time-series variance of returns
$$\sigma ^2_P$$ on the X-axis. For example, if we wanted to create the mean-variance parabola for one day in
the market, we could obtain the time-series variance of returns (X-axis) by using returns for an asset computed
every 10 minutes during the trading day. The mean 10-minute return and variance of returns for each asset are
plotted in the graph as a point. G is the minimum variance portfolio on the parabola with expected return
$$E(R_G)$$. Portfolios to the right of G on the upper half of the parabola lie on the efficient frontier (i.e.,
portfolios with the highest expected return unit total risk as measured by the variance of returns). Portfolios to the
right of G on the lower half of the parabola are inefficient minimum variance portfolios.

Fig. 10.1 Locating orthogonal portfolios $$I^*$$ and $$ZI^*$$ on the mean-variance parabola (Source Kolari et al. 2021, p. 59)

Before proceeding further, let’s review Black’s (1972) zero-beta CAPM. Dropping the time subscript t due to a
one-period model, the expected return for the ith asset is specified as follows:
$$\begin{aligned} E(R_i)=E(R_{ZI}) +\beta _{i,I}[E(R_I)-E(R_{ZI})], \end{aligned}$$ (10.1)
where $$E(R_I)$$ is the expected return on an efficient portfolio on the upper boundary of the parabola,
$$E(R_{ZI})$$ is the expected return on an orthogonal zero-beta portfolio on the lower boundary of the
parabola, and $$\beta _{i,I}$$ is beta risk associated with excess return of efficient portfolio I. The expected
return of the zero-beta portfolio replaces the riskless rate $$R_f$$ in the CAPM as the borrowing rate.
Unlike the CAPM, that requires the use of the market portfolio M on the efficient frontier, Black’s zero-beta
CAPM allows the use of any pair of efficient/inefficient minimum variance portfolios on the parabola that are
orthogonal to one another. If M could somehow be located, then ZM would be the zero-beta portfolio.
Returning to Fig. 10.1, suppose that the parabola is a picture based on asset returns for one day in the market.
In their book, KLH mathematically proved that the width or span of the parabola is largely determined by the
cross-sectional standard deviation (or variance) of returns for all assets denoted $$\sigma ^2_{at}$$ in a
given period of time t (e.g., one day).2 This previously unknown result makes sense. The opportunity set of all the
assets in the market is plotted within the parabola. The Y-axis shows the distribution of asset returns in the market.
This distribution has a mean and variance of returns. The variance is obviously the width of the parabola, such that
the mean market return must logically be approximately in the middle of the distribution. They denoted the mean
market return at time t as $$E(R_{at})$$. In another result not previously recognized, the mean market
return is proposed to be approximately equal to the expected return on the minimum variance portfolio
$$E(R_{Gt})$$, which splits the parabola into two halves on the axis of symmetry shown as the dashed line
in Fig. 10.1. The simple logic is that the mean market return $$E(R_{at})$$ lies approximately in the middle
of the distribution of returns on the Y-axis.
Most researchers prior to these findings believed that the mean market return (e.g., the value-weighted
CRSP index or S &P 500 index) was located somewhere in the vicinity of the efficient frontier. However,
according to KLH, the mean market return is far below the efficient frontier. As you will recall in previous
chapters, researchers have tested the CAPM with U.S. stock returns using the value-weighted CRSP index as a
proxy for the efficient market portfolio M. If this index lies somewhere along the axis of symmetry, it obviously is
far below the true market portfolio M. This mean market return location helps to explain empirical tests of the
CAPM that find a flatter relation between expected returns and beta risk estimates than predicted by the CAPM.
This flatter relation would result in a higher intercept or $$\alpha$$ term than otherwise. Consistent with the
Roll (1977) critique, an inefficient estimate of the market portfolio cannot be used to test the CAPM. If one accepts
these new insights about the mean-variance parabola, then researchers have erroneously rejected the CAPM by
using an inefficient market index m as a proxy for M.
Using the geometry shown in Fig. 10.1, the authors located two orthogonal portfolios denoted $$I^*$$
and $$ZI^*$$ on the parabola. Symmetric rays extending from $$E(R_G) \approx E(R_a)$$ yield two
tangent points to the minimum variance boundary of the parabola. Portfolio $$I^*$$ lies on the efficient
frontier, and portfolio $$ZI^*$$ is on the inefficient, minimum variance boundary of the parabola. This
orthogonal pair of portfolios is unique due to having the same time-series variance of returns and therefore are
marked with $$*$$ superscripts as $$\sigma ^2_{I^*}=\sigma ^2_{ZI^*}$$. Since the ZCAPM is a
one-period model, as in Black’s zero-beta CAPM, the time subscript t is dropped. These two portfolios are
important because their expected returns can be defined as follows:
$$\begin{aligned} E(R_{I^*})\approx &amp; {} E(R_{a})+\sigma _{a}\end{aligned}$$ (10.2)

$$\begin{aligned} E(R_{ZI^*})\approx &amp; {} E(R_{a})-\sigma _{a}. \end{aligned}$$ (10.3)

That is, to reach these portfolios in Fig. 10.1, an investor needs to move along the axis of symmetry at
$$E(R_{a})$$ to the right of portfolio G until they reach the point at which the time-series variance of
portfolios $$I^*$$ and $$ZI^*$$ are equal to one another, or
$$\sigma ^2_{I^*}=\sigma ^2_{ZI^*}$$. From this point, moving up or down by the cross-sectional
variance of returns (return dispersion) $$\sigma ^2_a$$ can be used to locate portfolios $$I^*$$ and
, respectively.3 Notice here that the geometry of the ZCAPM is much different from the CAPM.
$$ZI^*$$
The CAPM locates market portfolio M by means of a ray from the riskless rate $$R_f$$ to a tangent point
with the efficient frontier of the parabola. The ZCAPM moves horizontally along the axis of symmetry and then
vertically up or down to locate orthogonal portfolios on the efficient/inefficient boundaries of the parabola.
Next, they substitute the definitions of $$E(R_{I^*})$$ and $$E(R_{ZI^*})$$ in Eqs. (10.2) and
(10.3), respectively, into zero-beta CAPM Eq. (10.1) as follows:
$$\begin{aligned} E(R_i)= &amp; {} E(R_{ZI^*})+\beta _{i,I^*}[E(R_{I^*})-E(R_{ZI^*})]\nonumber
\\= &amp; {} E(R_{a})-\sigma _{a}+\beta _{i,I^*} \{[E({R}_{a})+\sigma _{a}]-[E({R}_{a})-\sigma
(10.4)
_{a}]\}\nonumber \\= &amp; {} E({R}_{a})+(2\beta _{i,I^*}-1)\sigma _{a}\nonumber \\ E({R}_i)= &amp;
{} E({R}_{a}) + Z_{i,a}^{*}\sigma _{a}, \end{aligned}$$
where $$Z_{i,a}^{*} = 2\beta _{i,I^*}-1$$. This relation is the theoretical ZCAPM in the absence of a
riskless asset.
Assuming a riskless asset f exists with rate $$R_f$$, they subsequently consider three assets I, ZI, and f.
In this three-asset case, the expected return of the ith asset can be defined as:
$$\begin{aligned} E({R}_i)= w_{I^*}E({R}_{I^*})+w_{ZI^*}E({R}_{ZI^*})+w_{f}R_f.
(10.5)
\end{aligned}$$
This equation can be re-written into the initial form of the theoretical ZCAPM:
$$\begin{aligned} E({R}_i)=(w_{I^*}+w_{ZI^*})E({R}_{a})+(w_{I^*}-w_{ZI^*})\sigma
(10.6)
_{a}+w_{f}R_f, \end{aligned}$$
where $$w_{I^*}$$, $$w_{ZI^*}$$, and $$w_{f}$$ are asset weights that sum to one. It is
assumed that investors can hold long and short positions in these assets. Rearranging terms, we get:
$$\begin{aligned} E({R}_i)-R_f=(w_{I^*}+w_{ZI^*})[E({R}_{a})-R_f]+(w_{I^*}-w_{ZI^*})\sigma
(10.7)
_{a}. \end{aligned}$$
Finally, using Eq. (10.4), this equation can be re-written in the final form of the theoretical ZCAPM:
$$\begin{aligned} E({R}_i)-R_f=\beta _{i,a}[E({R}_{a})-R_f] + Z_{i,a}^*\sigma _{a}, \end{aligned}$$ (10.8)
where $$\beta _{i,a}=w_{I^*}+w_{ZI^*}$$ is the beta risk coefficient measuring the sensitivity of the ith
asset’s excess returns to average market excess returns of all assets, and $$Z_{i,a}^*=w_{I^*}-w_{ZI^*}$$
is the zeta risk coefficient measuring the sensitivity of an asset’s excess returns to the market return dispersion of
all assets.4 Because both the average market return $$E(R_a)$$ and market return dispersion
$$\sigma _a$$ are observable in the real world,5 the problem of finding proxies in the real world is greatly
reduced. Again, good proxies m for the expected return on the market portfolio M in the CAPM or zero-beta
portfolio ZM in the zero-beta CAPM are difficult to construct and subject to controversy. Thus, the ZCAPM has
the advantage of estimable proxies of its two factors.
It is important to recognize that beta risk in the ZCAPM is different from the CAPM. In the CAPM, beta risk is
associated with the expected return on efficient market portfolio M, which is very difficult to identify or locate in
the real world. However, in the ZCAPM, beta risk is related to the average return on the portfolio of all assets a in
the market (not M) that is not efficient. Additionally, the zeta risk coefficient $$Z_{i,a}^*$$ can have a
positive or negative sign to take into account both portfolios $$I^*I^*$$ and $$ZI^*$$ on the
boundary of the parabola. As pointed out by Kolari et al. (2021, p. 75):

This two-sided, opposite market volatility risk is a distinctive feature of the ZCAPM that arises by virtue of
the mean-variance investment parabola with span determined by return dispersion.

Fig. 10.2 The ZCAPM with beta risk related to average excess market returns and zeta risk associated with positive and negative sensitivity to return
dispersion (Source Kolari et al. 2021, p. 72)

Figure 10.2 provides an illustration of the ZCAPM. The beta risk of the ith asset is based on the relation
between the ith asset’s excess returns, or $$R_{i}-R_{f}$$ on the Y-axis, and average excess market returns,
or $$R_{a}-R_{f}$$ on the X-axis. Zeta risk is determined by the relation between the asset’s excess
returns, or $$R_{i}-R_{f}$$, and market return dispersion, or $$\sigma _a$$. Here we see the two-
sided, opposite effects of zeta risk on asset returns.

Fig. 10.3 Two-sided, opposite effects of return dispersion on the expected returns of two assets (Source Kolari et al. 2021, p. 94)

Figure 10.3 gives another more simplified view of two-sided return dispersion effects. Assume there are two
time periods t = 1 and 2 in addition to two assets B and C. Return dispersion in the market increases from
$$\sigma _{a1}=1$$% at time 1 to 2% at time 2. Due to this increase in cross-sectional market volatility
from time 1 to 2, asset C experiences increasing returns but the returns of asset B decrease at time 2. Of course, if
return dispersion decreased over time, asset C would experience decreasing returns and asset B increasing returns.
Thus, return dispersion has opposite effects on the returns of assets depending on where they are located in the
distribution of returns.
In combination with beta risk, the dual systematic risk effects of return dispersion in the market give the mean-
variance parabola its upper and lower boundaries. We can think of the upper (lower) boundary as having some
amount of positive (negative) zeta risk—for example, $$Z_{i,a}^* = 0.8\ (-0.8)$$. Along the upper (lower)
boundary, assets will have increasing levels of beta risk. Hence, beta and zeta risk are important in defining the
architecture of the mean-variance parabola.
As a visual example, Fig. 10.4 shows some results for 25 U.S. stock portfolios based on the ZCAPM in Kolari
et al. (2021).6 The top horizontal boundary (or frontier) contains the five stock portfolios with the highest average
zeta risk. The frontier has an upward slope due to increasing levels of beta risk among these five portfolios. The
bottom boundary contains five stock portfolios with the lowest zeta risk in addition to varying levels of beta risk
from low to high. At each beta risk level, there is a vertical beta risk curve (i.e., represented by a dashed line) that
intersects different zeta risk curves. Taken together, the intersecting latticework of beta-zeta risk portfolios in the
graph takes on the general form of a mean-variance parabola. In effect, stock portfolios within the parabola have
beta and zeta risk characteristics that jointly determine their expected returns. Rather than being randomly
distributed within the parabola, portfolios are ordered by their systematic risks related to average market
returns and market return dispersion. Hence, equilibrium asset prices are established within the parabola due to
these two market forces.

Fig. 10.4 This graph shows the average beta risk, zeta risk, and one-month-ahead equal-weighted returns for 25 beta-zeta sorted portfolios. These long-
only portfolios approximate the shape on an investment parabola. In each one-year estimation window, the empirical ZCAPM is estimated using daily
returns to obtain the beta and zeta risk parameters. The analysis period is January 1965–December 2018 (Source Kolari et al. 2021, p. 272)
10.2 Empirical ZCAPM
How can we model the positive and negative effects of return dispersion on real-world asset returns? Kolari, Liu,
and Zhang proposed the following novel empirical ZCAPM:
$$\begin{aligned} {R}_{it}-R_{ft}=\alpha _i+\beta _{i}({R}_{at}-R_{ft})+Z_{i}D_{it}{\sigma }_{at}+
(10.9)
{u}_{it}, ~~t = 1,\ldots , T \end{aligned}$$
where $$Z_{i}$$ measures sensitivity to return dispersion $$\sigma _{at}$$, $$D_{it}$$ is a signal variable
with values $$+1$$ and $$-1$$ representing positive and negative return dispersion effects on stock returns at
time t, respectively, $$u_{it} \thicksim \text {iid N}(0,\sigma _i^2)$$, and other notation is as before. The
dummy signal variable $$D_{it}$$ is not observable and therefore is considered to be an unknown or latent
(hidden) variable. Unfortunately, due to the hidden variable $$D_{it}$$, commonly used ordinary least squares
(OLS) regression cannot be used to estimate the empirical ZCAPM.
To take into account signal variable $$D_{it}$$, the well-known statistical method of expectation-
maximization (EM) regression can be employed.7 In EM, hidden variable $$D_{it}$$ is defined as an
independent random variable with the following two-point distribution:
$$\begin{aligned} D_{it}={\left\{ \begin{array}{ll} +1 &amp;{} \text {with probability}~ p_i\\ -1
(10.10)
&amp;{} \text {with probability}~ 1-p_i, \end{array}\right. } \end{aligned}$$
where $$p_i$$ (or $$1-p_i$$) is the probability of a positive (or negative) return dispersion effect, and
$$D_{it}$$ is independent of $$u_{it}$$. No previous studies utilize EM regression in an asset pricing
model.
Notice that, in the empirical ZCAPM relation (10.9), the interaction term $$Z_{i,a}D_{i,t}$$ has two
possible values equal to $$+Z_{i,a}$$ or $$-Z_{i,a}$$ based on the sign of signal variable
$$D_{i,t}$$ . Since the mean of binary signal variable $$D_{i,t}$$ equals $$2 p_i -1$$, we can
define $$Z_{i,a}^* = Z_{i,a} (2p_i-1)$$, which results in the final form of the empirical ZCAPM:
$$\begin{aligned} R_{it}-R_{ft}=\beta _{i,a}(R_{at}-R_{ft}) + Z_{i,a}^* \sigma _{at}+u_{it}, ~~t =
(10.11)
1,\ldots , T. \end{aligned}$$
EM regression provides estimates of beta risk coefficient $$\beta _{i,a}$$ as well as zeta risk coefficient
$$Z_{i,a}^*$$ that coincides with theoretical ZCAPM relation (10.8). The positive or negative sign of
$$Z_{i,a}^*$$ is determined by the probability $$p_i$$ of signal variable $$D_{it}$$ in sample
period $$t = 1,\ldots , T$$ . If $$p_i &gt; 1/2 \text { (or } &lt;1/2)$$ , $$Z_{i,a}^*$$ will have a
positive (or negative) sign. In words, $$Z_{i,a}^*$$ gives the average increase or decrease of asset returns in
response to a one-unit change in return dispersion $$\sigma _{at}$$.
Unlike other empirical asset pricing models, the empirical ZCAPM does not have an intercept (
$$\alpha$$) term. The EM algorithm does not utilize an intercept to minimize squared error terms as in OLS
regression. For this reason, KLH repeated cross-sectional tests discussed in the next section with and without
intercepts in time-series multifactor models. They found that the cross-sectional tests of factor prices were
unaffected for the most part by time-series regressions that set the intercept to zero. However, no Gibbons et al.
(1989) (GRS) tests of the joint equality of $$\alpha$$s are possible for the empirical ZCAPM to compare to
other models.
For interested readers, KLH have placed Matlab, R, and Python programs for EM estimation of the empirical
ZCAPM on GitHub (https://github.com/zcapm). Their R programs estimate the ZCAPM faster than the Matlab and
Python programs. Programs for running Fama–MacBeth cross-sectional regression tests are provided on the
GitHub website also.

10.3 Empirical Evidence

Let’s look at some of the empirical evidence on the ZCAPM in KLH’s book. They employed daily U.S. stock
returns from the period 1964 to 2018. The value-weighted CRSP index return was used to proxy $$R_{at}$$ on
each day t. The value-weighted market return dispersion on each day t was computed as follows:
$$\begin{aligned} \sigma _{at} = \sqrt{\frac{n}{n-1}\displaystyle \sum \limits _{i=1}^n w_{it-1} (10.12)
(R_{it} - R_{at})^2}, \end{aligned}$$
where n is the total number of stocks, $$w_{it-1}$$ is the previous day’s market value weight for the ith stock
(i.e., market capitalization of the stock divided by the market capitalization of all n stocks), $$R_{it}$$ is the
return of the ith stock on day t, and $$R_{at}$$ is value-weighted average return of all available stocks in the
CRSP database on day t. In their sample period, the average daily market return was 0.04% (i.e., 0.88% per month)
compared to 0.99% for daily return dispersion. These descriptive statistics indicate that return dispersion is quite
large compared to mean market daily returns. Also, because there was almost zero correlation between these two-
asset pricing factors, the factors $$R_{at}$$ and $$\sigma _{at}$$ are orthogonal to one another for the most
part and therefore contain different market information.
The authors conducted out-of-sample (or one-month-ahead), cross-sectional Fama and MacBeth (1973) tests
for a number of different test asset portfolios. Empirical ZCAPM relation (10.11) was estimated using one year of
daily returns. The U.S. Treasury bill rate proxied the riskless rate. Following standard practices, Fama–MacBeth
analyses were rolled forward one month at a time to generate a series of estimated market prices of beta and zeta
risk loadings (denoted $$\lambda _a$$ and $$\lambda _{RD}$$, respectively). To comparatively
evaluate the performance of the ZCAPM, they ran tests for the CAPM, Carhart (1997) four-factor model, and
Fama and French (1992, 1993, 2015, 2018) three-, five-, and six-factor models. In forthcoming empirical evidence,
we show results for not only the ZCAPM but these popular single and multifactor models covered in previous
chapters. Also, results are compared to often-cited four-factor models by Hou et al. (2015) and Stambaugh and
Yuan (2017).

10.3.1 Risk and Return Relations

To begin their analyses, KLH reported evidence on the relation between beta risk and zeta risk with respect to one-
month-ahead stock returns. Using the 97 portfolios discussed above, the empirical ZCAPM in Eq. (10.11) was run
using daily returns for one year to obtain $$\beta _{i,a}$$ and $$Z_{i,a}^*$$ estimates for each portfolio. The
actual returns $$R_{it+1}$$ in the next one month were retained. The process was repeated by rolling forward
one month to get new estimates of $$\beta _{i,a}$$ and $$Z_{i,a}^*$$ as well as the next one-month returns
$$R_{it+1}$$ for portfolios. Repeating this process from 1964 to the end of 2018, they generated a series of 648
monthly values for $$\beta _{i,a}$$, $$Z_{i,a}^*$$, and $$R_{it+1}$$. Next, the average values of these
values were computed for each portfolio.

Fig. 10.5 Out-of-sample cross-sectional ZCAPM relationship between average one-month-ahead realized excess returns in percent (Y-axis) and average
beta risk $$\beta _{i,a}$$ in the previous 12-month estimation period (X-axis). Results are shown for 25 size-BM portfolios used in the Fama and
French studies. Portfolios are sorted into zeta risk $$Z^*_{i,a}$$ quintiles and then beta risk $$\beta _{i,a}$$ quintiles within each
$$Z^*_{i,a}$$ quintile. The analysis period is January 1965–December 2018 (Source Kolari et al. 2021, p. 141)

For test asset portfolios, they used 25 size and BM sorted portfolios from Kenneth French’s data website.8
These portfolios are often used by Fama and French as well as other researchers in asset pricing model tests.
Figures 10.5 and 10.6 graphically show the relation between average one-month-ahead excess returns (Y-axis) and
average beta and zeta risk coefficients (X-axis), respectively. It is apparent from Fig. 10.5 that returns are not
related to beta risk. Most of the portfolios have betas near one, which is the beta of the CRSP market risk.
Moreover, a slightly inverse relation can be seen, which is opposite of CAPM theory. Concerning these results, as
discussed earlier, beta risk is different in the ZCAPM than CAPM.

Fig. 10.6 Out-of-sample cross-sectional ZCAPM relationship between average one-month-ahead realized excess returns in percent (Y-axis) and average
zeta risk $$Z^*_{i,a}$$ in the previous 12-month estimation period (X-axis). Results are shown for 25 size-BM portfolios used in the Fama and
French studies. Portfolios are sorted into beta risk $$\beta _{i,a}$$ quintiles and then zeta risk $$Z^*_{i,a}$$ quintiles within each
$$\beta _{i,a}$$ quintile. The analysis period is January 1965–December 2018 (Source Kolari et al. 2021, p. 140)

In stark contrast, Fig. 10.6 reveals a very strong relation between returns and zeta risk. Given zeta risks for the
25 portfolios, their future returns in the next month line up almost perfectly with the order of their previously
estimated risks. This finding is extraordinary. Additionally, KLH sorted the 25 portfolios into beta risk quintiles
and then zeta risk quintiles within each beta risk quintile. This sorting reveals that small stocks tended to have
larger betas than big stocks. This pattern occurs in Fig. 10.5 also.
10.3.2 Predicted and Actual Return Relations
Are predicted returns using the ZCAPM related to future actual returns? Recall from Chap. 4 that Fama and
MacBeth (1973) argued that normative models, which are intended to help investors make better decisions, are
only valid to the extent that past information is related to future returns. To evaluate predicted versus actual
returns, we estimated beta and zeta risk parameters using the empirical ZCAPM with one year of daily returns. In
the next month on an out-of-sample basis, estimated beta and zeta coefficients were multiplied by average excess
market returns $$R_{at}-R_{ft}$$ and market return dispersion $$\sigma _{at}$$ on each day t to compute
predicted (or fitted) excess returns for each portfolio. These daily predicted excess returns are used to compute a
one-month predicted excess return in the next month. The realized (or actual) excess return in the next out-of-
sample month is recorded also. This process is rolled forward month-by-month until December 2018 to get out-of-
sample monthly series of predicted and realized excess returns for each portfolio. Lastly, averages of these excess
returns from January 1965 to December 2018 are computed. In forthcoming Table 10.1, we utilize these average
excess returns to assess the goodness-of-fit of different models. To do this, we regress average realized excess
returns on average predicted excess returns for each model to estimate their $$R^2$$ value.

Fig. 10.7 Out-of-sample cross-sectional relationship between average one-month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25 size-BM sorted portfolios: Fama and French three-factor model in Panel A and
empirical ZCAPM in Panel B. The analysis period is January 1965 to December 2018 (Source Kolari et al. 2021, p. 146)

Using the 25 size and BM sorted portfolios, Fig. 10.7 shows the results for the Fama and French three-factor
model versus the empirical ZCAPM. The three-factor model performs well in these tests. The (out-of-sample)
predicted and realized excess returns in these out-of-sample analyses line up closely to the 45-degree line. By
comparison, the ZCAPM performs even better. Notice that the three-factor model has one stray portfolio, which is
the small-size portfolio with low BM (growth).9 However, this stray portfolio is priced by the ZCAPM as it lies
exactly on the 45-degree line.

Fig. 10.8 Out-of-sample cross-sectional relationship between average one-month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25 size-BM sorted plus 47 industry portfolios: Fama and French Fama and French
three-factor model in Panel A and empirical ZCAPM in Panel B. The analysis period is January 1965–December 2018 (Source Kolari et al. 2021, p. 147)

Figure 10.8 repeats these experiments with the 25 size and BM portfolios plus 47 industry portfolios. Here we
see that the three-factor model is not useful, but the ZCAPM does a fairly good job. Industries are exogenous to
the three-factor model, unlike the endogenous size and BM portfolios (i.e., as discussed in Chapter 8, the size and
value factors are related to the size and BM portfolios). In this regard, referring back to Fig. 10.6, the size and BM
portfolios are exogenous to the ZCAPM. Even so, the ZCAPM outperforms the three-factor model. Together, these
results suggest that the ZCAPM has better goodness-of-fit than the three-factor model. The tables have turned!
Goodness-of-fit is higher in the CAPM model than the multifactor model.

Fig. 10.9 Out-of-sample cross-sectional relationship between average one-month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25 profit-investment sorted portfolios: Fama and French six-factor model in Panel A
and empirical ZCAPM in Panel B. The analysis period is January 1965 to December 2018 (Source Kolari et al. 2021, p. 150)

Fig. 10.10 Out-of-sample cross-sectional relationship between average one-month-ahead realized excess returns in percent (Y-axis) and average one-
month-ahead predicted (fitted) excess returns in percent (X-axis) for 25 profit-investment sorted plus 47 industry portfolios: Fama and French Fama and
French six-factor model in Panel A and empirical ZCAPM in Panel B. The analysis period is January 1965–December 2018 (Source Kolari et al. 2021, p.
151)

Figures 10.9 and 10.10 repeat these analyses using the Fama and French six-factor model. Here the 25 size and
BM portfolios are replaced with 25 profit-investment sorted portfolios from Kenneth French’s data website.
Clearly, the six-factor model performs quite well and does better than the three-factor model, but the ZCAPM still
outperforms it. Also, the six-factor model has difficulty in fitting returns for industry portfolios. When test asset
portfolios are used that are not related to long/short factors, the multifactor models are more likely to fall down.

10.3.3 Fama and MacBeth Cross-Sectional Tests

Out-of-sample Fama and MacBeth (1973) cross-sectional tests for the CAPM, ZCAPM, and a number of
discretionary multifactor models are reported in Table 10.1. These selected results were taken from Kolari et al.
(2021) and are representative of the many test results reported in their book. The ZCAPM is benchmarked against
the CAPM and popular Carhart four-factor model as well as Fama and French three-, five- and six-factor models.
In step one, the ZCAPM and other models are estimated with one year of daily returns beginning in 1964. In step
two, one-month-ahead (out-of-sample) returns in January 1965 are regressed on previously estimated risk
parameters in step one. The process is rolled forward one month at a time and repeated until the one-month-ahead
return in December 2018. A monthly series of 648 estimated market prices of risk, denoted $$\lambda _k$$, for
the kth factor and 648 one-month-ahead returns are generated for each test asset portfolio. Averages of
$$\lambda _k$$s are shown in the table.
Table 10.1 Out-of-sample Fama–MacBeth cross-sectional tests for ZCAPM regression factor loadings compared to other asset pricing models in the
period January 1965–December 2018: 12-month rolling windows

Model $$\hat{\al $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb

pha }$$ da }_m$$ da }_{ RD }$$ da }_{ SMB da }_{ HML da }_{ MOM da }_{ RMW da }_{ CMA 2$
}$$ }$$ }$$ }$$ }$$
Panel A: 25 size-BM sorted portfolios
CAPM 0.98 −0.35 0.4
(4.03) (−0.35)
ZCAPM 0.78 −0.19 0.46 0.9
(3.26) (−0.77) (4.22)
Three- 0.93 −0.41 0.17 0.31 0.6
factor (4.98) (−1.96) (1.27) (2.64)
Four- 1.12 −0.60 0.16 0.29 0.08 0.6
factor (5.84) (−2.83) (1.29) (2.57) (0.35)
Five- 0.98 −0.45 0.21 0.28 0.13 0.11 0.7
factor (4.94) (−2.06) (1.73) (2.49) (1.10) (0.84)
Six- 1.07 −0.55 0.22 0.30 0.16 0.12 0.06 0.8
factor (5.39) (−2.52) (1.80) (2.69) (0.68) (1.03) (0.45)
Panel B: 47 industry portfolios
CAPM 0.57 0.03 0.0
(2.95) (0.15)
ZCAPM 0.49 0.11 0.32 0.7
(2.71) (0.53) (4.07)
Three- 0.43 0.11 0.02 0.14 0.
factor (2.21) (0.48) (0.14) (1.17)
Four- 0.46 0.08 0.04 0.08 0.43 0.3
factor (2.31) (0.37) (0.31) (0.73) (2.09)
Five- 0.36 0.19 0.07 0.09 0.20 0.24 0.
factor (1.78) (0.83) (0.51) (0.76) (2.03) (2.48)
Six- 0.51 0.04 0.04 0.06 0.39 0.20 0.17 0.5
factor (2.53) (0.16) (0.32) (0.52) (1.92) (2.05) (1.78) (1
Panel C: 25 beta-zeta sorted portfolios
CAPM 0.49 −0.06 0.
(3.23) (−0.26)
ZCAPM 0.29 0.17 0.35 0.9
(1.91) (0.83) (4.03)
Three- 0.66 −0.05 −0.28 0.14 0.6
factor (3.34) (−0.24) (−1.49) (1.00)
Four- 0.54 −0.03 −0.25 0.25 0.28 0.7
factor (3.50) (−0.15) (−1.51) (1.76) (1.61)
Five- 0.52 −0.03 −0.19 0.25 0.17 0.09 0.6
factor (3.06) (−0.13) (−1.11) (1.56) (1.18) (0.77)
Model $$\hat{\al $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb $$\hat{\lamb
pha }$$ da }_m$$ da }_{ RD }$$ da }_{ SMB da }_{ HML da }_{ MOM da }_{ RMW da }_{ CMA 2$
}$$ }$$ }$$ }$$ }$$
Six- 0.62 −0.11 −0.14 0.30 0.26 0.17 0.02 0.7
factor (3.86) (−0.54) (−0.88) (1.90) (1.50) (1.30) (0.14)
Panel D: 97 combined portfolios
CAPM 0.64 −0.06 0.0
(3.93) (−0.31)
ZCAPM 0.48 0.09 0.38 0.8
(3.24) (0.47) (5.42)
Three- 0.52 0.002 0.06 0.18 0.0
factor (3.55) (0.01) (0.49) (1.64)
Four- 0.53 −0.01 0.09 0.17 0.43 0.2
factor (3.75) (−0.06) (0.75) (1.64) (2.54)
Five- 0.44 0.09 0.12 0.14 0.21 0.21 0.2
factor (2.97) (0.45) (1.02) (1.35) (1.99) (2.36)
Six- 0.57 −0.04 0.14 0.17 0.40 0.19 0.16 0.3
factor (4.02) (−0.24) (1.20) (1.68) (2.40) (1.94) (1.84)

This table reports selected results from Kolari et al. (2021). Out-of-sample (one-month-ahead) estimated prices of
risk were estimated using the standard two-step Fama-MacBeth cross-sectional tests. Estimated prices of risk are
denoted $$\hat{\lambda }_k$$ for the kth factor in monthly percent return terms (t-statistics in parentheses).
Factors are denoted as m (CRSP index, see footnote), RD (return dispersion), $$\textit{SMB}$$ (size),
$$\textit{HML}$$ (value), $$\textit{MOM}$$ (momentum), $$\textit{RMW}$$ (profit), and
$$\textit{CMA}$$ (capital investment). Value-weighted returns were used in the period January 1965–
December 2018. Different sets of test asset portfolios were employed as shown in Panels A–D. The results
shown below are based on Kolari et al. (2021, Table 7.1, p. 166)
aIn the ZCAPM, the price of beta risk associated with CRSP index excess returns is denoted

$$\hat{\lambda }_a$$ rather than $$\hat{\lambda }_m$$ in the other models

The first notable finding in Table 10.1 is the relatively high $$R^2$$ values of the ZCAPM compared to
the other models. In Panels A and C, the ZCAPM achieves $$R^2$$ estimates of 94% and 96%,
respectively. This goodness-of-fit is near perfect! Almost all variation in cross-sectional stock returns in these
portfolios is explained by the ZCAPM. The near perfect fit of the ZCAPM is remarkable in light of fact that these
tests utilize one-month-ahead (out-of-sample) returns. The CAPM has fairly good fit at 48% for the 25 size and
BM sorted portfolios (Panel A) but is zero for 47 industry portfolios (Panel B). The highest multifactor model
$$R^2$$ value of 80% is achieved by the six-factor model in the 25 size and BM portfolios (Panel A) but
falls to 50% for 47 industry portfolios (Panel B) and only 35% for 97 combined portfolios (Panel D). In the latter
two portfolio tests, the ZCAPM posted respectable 70% and 83% $$R^2$$ values. These results suggest that
the ZCAPM outperforms the CAPM and multifactor models in general and by large margins with industry
portfolios in particular.
Looking at the market prices of risk, or $$\lambda _k$$s, in Table 10.1, the market price of return
dispersion ( $$\lambda _{RD}$$ ) has the highest t-values of statistical significance ranging from 4.03 to 5.42
in the different panels. Size factor loadings ( $$\lambda _{SMB}$$) never reach statistical significance at the
5% level. Loadings for other factors, including value ( $$\lambda _{HML}$$), momentum (
$$\lambda _{MOM}$$), profit ( $$\lambda _{RMW}$$), and capital investment (
$$\lambda _{CMA}$$), are intermittently significant in different models and test asset portfolios. And,
market factor loadings are normally insignificant but are significant with a negative sign for the 25 size and BM
portfolios in some models (opposite of CAPM theory).
Notice that none of the multifactor loadings achieves a t-statistic greater than 3. In extensive tests of over 300
multifactors, Harvey et al. (2016) and Chordia et al. (2020) documented evidence that most multifactors are false
discoveries.10 For this reason, these authors recommended that asset pricing factors should surpass a t-statistic
threshold of 3 to be considered significant in cross-sectional tests. This hurdle is relatively high. Rarely are widely
used factors able to exceed this threshold, especially in out-of-sample cross-sectional tests. Using this criterion,
only zeta risk loadings in the ZCAPM pass recent standards for the validity of factors. Other models are subject to
question in view of these findings.
Overall, the results in Table 10.1 show that the empirical ZCAPM outperforms the CAPM and popular
multifactor models in standard Fama and MacBeth cross-sectional tests. KLH also conducted cross-sectional tests
of the Hou et al. (2015) q-factor model with four factors (viz., market, equity capitalization, investment to assets
ratio, and return on equity) as well as the Stambaugh and Yuan (2017) mispricing model with four factors (market,
size, management, and performance). None of the t-statistics for estimated $$\lambda _k$$s exceeded three.
Moreover, the goodness-of-fit was less than the ZCAPM, especially for industry portfolios as in the tests of other
models in Table 10.1. Hence, using similar test assets and sample periods, the ZCAPM outperformed these four-
factor models in cross-sectional tests.
More recent work by Kolari et al. (2022b) extended the U.S. stock return analyses in KLH to a longer sample
period of 1927 to 2020. The results corroborated those in KLH. Also, Kolari et al. (2022a) extended the analyses to
Canada, Japan, Germany, and Japan. Again the results confirmed findings in KLH—namely, the ZCAPM
outperformed multifactor models especially when using exogenous industry portfolios.
A natural question is: Why does the ZCAPM consistently outperform multifactor asset pricing models? As
discussed in KLH and mentioned earlier, the long/short multifactors are themselves rough measures of return
dispersion. They capture different slices of the total return dispersion as measured by $$\sigma _{at}$$. At
times the market prices of multifactor loadings can turn negative (e.g., see the 25 beta-zeta portfolios in Panel C).
These negative prices are difficult to explain in the context of the multifactor models. However, in line with the
ZCAPM, they are picking up negative zeta risk. As more factors are added to multifactor models, they come closer
to capturing total return dispersion. Of course, the multifactors can change over time and shift their position within
the distribution of total returns. In general, all multifactors are subsumed or contained within total return
dispersion.
Fama and French and others loosely link multifactor models to the hedge portfolio theories of Merton’s (1973)
intertemporal CAPM and Ross’ (1976) arbitrage pricing theory (APT). Further linking theory to empirical models,
KLH assert that multifactor models are closely linked to total return dispersion in the ZCAPM. Viewed as return
dispersion measures, multifactors are therefore related to both KLH’s ZCAPM and Black’s (1972) zero-beta
CAPM in addition to the CAPM of Treynor (1961, 1962), Sharpe (1964), Lintner (1965), and Mossin (1966). In
a very real sense, all of these models are cousins of the CAPM! As already discussed, the main complaint against
the CAPM by Fama and French and others was its poor empirical fit to stock returns. The ZCAPM overcomes this
objection by dramatically improving the empirical fit of a CAPM-based model.

10.4 Summary
Kolari, Liu, and Huang (KLH) (2021) recently proposed a new CAPM dubbed the ZCAPM. The ZCAPM has two
factors: mean excess market returns and the cross-sectional standard deviation of returns for all assets in the
market or simply return dispersion. Return sensitivity to these two factors is measured by beta risk and zeta risk,
respectively. The ZCAPM is grounded in the Markowitz (1959) mean-variance investment parabola, equilibrium
conditions of the famous CAPM by Treynor (1961, 1962), Sharpe (1964), Lintner (1965), and Mossin (1966), and
the zero-beta CAPM of Black (1972).
KLH mathematically derived the ZCAPM as a special case of Black’s (1972) zero-beta CAPM. Two key
insights proposed by KLH are: (1) the width or span of the mean-variance parabola at time t is largely determined
by market return dispersion; and (2) the mean market return is approximately in the middle of the parabola
somewhere along its axis of symmetry. Using these insights, they identified two unique orthogonal portfolios on
the parabola denoted $$I^*I^*$$ and $$ZI^*$$ that have the same time-series variance of returns.
Upon substituting the formulas for the expected returns of these two portfolios into Black’s equation for the zero-
beta CAPM, they obtained a special case that represents the theoretical ZCAPM. Adding a riskless rate, they
derived the final form of the theoretical ZCAPM.
KLH specified a novel empirical ZCAPM with a dummy signal variable (equal to $$+1$$ or
$$-1$$) to capture either positive or negative return dispersion effects on asset returns. Since this signal
variable is unobservable, they used expectation-maximization (EM) regression methods to estimate the
probability that the signal variable is $$+1$$ or $$-1$$. Multiplying this probability times the zeta
risk regression coefficient, they obtained an estimate of zeta risk (denoted $$Z^*$$) for a stock. Beta
risk (denoted $$\beta$$) is the regression coefficient associated with mean excess market returns. The
empirical ZCAPM is a parsimonious two-factor model that employs factors that can be readily estimated from
available market data. There is no need to obtain a proxy for an efficient market portfolio M as in the CAPM or a
zero-beta portfolio ZM as in the zero-beta CAPM.
Empirical tests of the ZCAPM yielded impressive results. Using cross-sectional analyses, KLH showed that a
close relation exists between stock returns in the next month and previously estimated zeta risk coefficients. Also,
using step-ahead or out-of-sample returns in the next month, predicted (fitted) excess returns from the ZCAPM
were closer to realized (actual) excess returns in comparisons to the CAPM and Fama and French (1993, 2015,
2018) multifactor models.
More formal Fama and MacBeth (1973) cross-sectional tests using out-of-sample stock returns strongly
supported the ZCAPM over the CAPM and multifactor models. Comparative tests of the Carhart (1997) four-
factor model, Fama and French three-, five-, and six-factor models, Hou et al. (2015) four-factor q model, and
Stambaugh and Yuan (2017) four-factor model showed that:
(1)
estimated $$R^2$$ values from the empirical ZCAPM were noticeably higher on a consistent basis in tests
of different portfolios than those for the CAPM and multifactor models; and
(2)
zeta risk loadings estimated from the ZCAPM normally had t-statistics greater than 3, but multifactors in
other models did not exceed this recommended threshold.
A major difficulty of the CAPM as well as multifactor models is their poor pricing performance with respect to
exogenous industry portfolios. By contrast, the ZCAPM did a good job of pricing industry portfolios.
Why does the ZCAPM outperform popular multifactor models? The main reason is that multifactor models
depend on long/short portfolios as factors that are rough measures of return dispersion. A more complete and
accurate measure of return dispersion is the cross-sectional standard deviation of returns in the market. From this
perspective, all multifactor models are related to the ZCAPM and therefore the CAPM also. The main complaint of
poor empirical fit among CAPM models is overcome by the ZCAPM. Indeed, the ZCAPM consistently
outperforms the multifactor models in out-of-sample cross-sectional tests. Rather than the CAPM being
empirically dead, it is reborn in the form of the empirical ZCAPM. The history of asset pricing has come full
circle. The early theoretical models of the CAPM led to discretionary models with empirically justified
multifactors. However, the ZCAPM subsumes multifactors within its total return dispersion factor. Can machine
learning methods be developed to further enhance the empirical ZCAPM? Are there better proxies for mean
market returns and the cross-sectional return dispersion? By improving these proxies, the empirical ZCAPM could
be enhanced. It will be interesting to see what the future of asset pricing holds!

Questions
1.
Draw a picture of Markowitz’s mean-variance investment parabola. Label the Y-axis as $$E(R_P)$$ for
expected returns and X-axis as $$\sigma ^2_P$$ for the variance of returns. Where is the minimum
variance portfolio G? In your figure, show how Kolari, Liu, and Huang (KLH) (2021) located two orthogonal
portfolios with equal time-series variance of returns: $$I^*$$ on the efficient frontier, and $$ZI^*$$ on
the inefficient boundary of the parabola. According to KLH, what defines the width or span of the parabola?
Show this in your diagram along with the mean market return.
2.
In their picture of the mean-variance investment parabola, Kolari, Liu, and Huang (KLH) (2021) argued that
commonly used general stock market indexes (e.g., the value-weighted CRSP index or S &P 500 index) are
inefficient. Where do they locate these general indexes within the parabola? If this is true, what is the
implication to tests of the CAPM using the market model?
3.
Assuming a riskless asset exists, write the equation for the theoretical ZCAPM. What do beta risk and zeta
risk measure in this model? Does the ZCAPM have an empirical advantage in terms of testing over other
forms of the CAPM?
4. According to Kolari, Liu, and Huang (KLH) (2021), zeta risk (denoted by $$Z_{i,a}^*$$) in the ZCAPM
can be positive or negative in sign. Illustrate this two-sided zeta risk in a diagram with the ith asset’s excess
returns denoted $$R_{i}-R_{f}$$ on the Y-axis and average excess market returns denoted
$$R_{a}-R_{f}$$ on the X-axis. In your diagram, what is beta risk? Show how positive and negative zeta
risk affect the excess returns of assets.
5.
Show how two-sided return dispersion effects impact expected stock returns. Assume there are two time
periods t = 1 and 2 in addition to two assets B and C. Return dispersion in the market increases from
$$\sigma _{a1}=1$$% at time 1 to 2% at time 2. Due to this increase in cross-sectional market volatility
from time 1 to 2, show how expected asset returns of assets B and C are affected. What if return dispersion
decreased over time?
6.

Kolari, Liu, and Huang (KLH) (2021) argued that the mean-variance parabola is shaped by the dual
systematic risk effects of beta risk and zeta risk. How is the architecture of the parabola affected by these
systematic risks?
7.
Kolari, Liu, and Zhang (KLH) (2021) conducted out-of-sample (or one-month-ahead), cross-sectional Fama
and MacBeth (1973) tests of a number of different test asset portfolios. The final form of the empirical
ZCAPM was estimated using one year of daily returns. The U.S. Treasury bill rate proxied the riskless rate.
To begin their analyses, KLH reported evidence on the relation between beta risk and zeta risk with respect
to one-month-ahead stock returns. Using 97 stock portfolios, the empirical ZCAPM was run using daily
returns for one year to obtain $$\beta _{i,a}$$ and $$Z_{i,a}^*$$ estimates for each portfolio. The
actual returns $$R_{it+1}$$ in the next one month were retained. The process was repeated by rolling
forward one month to get new estimates of $$\beta _{i,a}$$ and $$Z_{i,a}^*$$ as well as the next one-
month returns $$R_{it+1}$$ for portfolios. Repeating this process from 1964 to the end of 2018, they
generated a series of 648 monthly values for $$\beta _{i,a}$$, $$Z_{i,a}^*$$, and $$R_{it+1}$$.
Next, the average values of these values were computed for each portfolio. What did they find?
8.
Following standard practices, Kolari, Liu, and Zhang (KLH) (2021) implemented Fama–MacBeth cross-
sectional regression tests. The analyses were rolled forward one month at a time to generate a series of
estimated market prices of beta and zeta risk loadings (denoted $$\lambda _a$$ and $$\lambda _{RD}$$
, respectively). To comparatively evaluate the performance of the ZCAPM, they ran tests for the
CAPM, Carhart (1997) four-factor model, and Fama and French (1992, 1993, 2015, 2018) three-, five-, and
six-factor models. For the often used 25 size and BM sorted portfolios from Kenneth French’s data website,
they produced graphs comparing realized excess returns to fitted excess returns for the three-factor model of
Fama and French (1992, 1993) versus the empirical ZCAPM. What did these graphs show? What about the
results for 47 industry portfolios? What do these results mean?
9.
Kolari, Liu, and Zhang (KLH) (2021) repeated the question above by comparing other multifactor models to
the empirical ZCAPM. For the six-factor model of Fama and French (2018) compared to the empirical
ZCAPM, using 25 profit-investment sorted portfolios, what did they find? For industry portfolios?
10.
In empirical tests of the ZCAPM, is there an explanation for why multifactor models were dominated by the
ZCAPM? Fama and French (1993, 1995, 2015, 2018) link that multifactor models to Merton’s (1973)
intertemporal CAPM and Ross’ (1976) arbitrage pricing theory (APT). Can the ZCAPM be linked to these
theoretical models?

Problems
1.
Given their new geometry of the parabola, how did Kolari, Liu, and Huang (KLH) (2021) define the expected
returns for the special case of orthogonal portfolios $$I^*$$ and $$ZI^*$$ on the parabola? Describe this
new geometry for locating these portfolios. How does this geometry differ from the CAPM and market
portfolio M.
2.
If we substitute the definitions of $$E(R_{I^*})$$ and $$E(R_{ZI^*})$$ into the zero-beta CAPM
Eq. 10.1, we can derive a theoretical ZCAPM without a riskless rate. Show how to get the theoretical ZCAPM
in this case.
3. Kolari, Liu, and Zhang (KLH) (2021) proposed a novel empirical ZCAPM to capture positive and negative
sensitivity to return dispersion (i.e., zeta risk). Write their empirical model. What is $$D_{it}$$ in this
 model? Using the definition of $$D_{it}$$, write the final form of the empirical ZCAPM. Why must the
expectation-maximization (EM) regression method must be used to estimate this final form?
4.

How did Kolari, Liu, and Zhang (KLH) (2021) compute mean market returns as well as the return dispersion
for U.S. stocks? Are equal-weighted or value-weighted return used in their computations of these factors?
5.
Kolari, Liu, and Zhang (KLH) (2021) conducted out-of-sample Fama and MacBeth (1973) cross-sectional
regression tests for the empirical ZCAPM and a number of other models, including the market model and
multifactor models. What did these tests reveal about the goodness-of-fit? About the significance of estimated
market prices of risk for different factors?

References
Black, F. 1972. Capital market equilibrium with restricted borrowing. Journal of Business 45: 444–454.
[Crossref]

Carhart, M.M. 1997. On persistence in mutual fund performance. Journal of Finance 52: 57–82.
[Crossref]

Chordia, T., A. Goyal, and A. Saretto. 2020. Anomalies and false rejections. Review of Financial Studies 33: 2134–2179.
[Crossref]

Dempster, A.P., N.M. Laird, and D.B. Rubin. 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical
Society 39: 1–38.

Fama, E.F., and K.R. French. 1992. The cross-section of expected stock returns. Journal of Finance 47: 427–465.
[Crossref]

Fama, F.F. and K.R. French 1995. Size and book-to-market factors in earnings and returns. Journal of Finance 50: 131–155.
[Crossref]

Fama, E.F., and K.R. French. 1993. The cross-section of expected returns. Journal of Financial Economics 33: 3–56.
[Crossref]

Fama, E.F., and K.R. French. 1996. The CAPM is wanted, dead or alive. Journal of Finance 51: 1947–1958.
[Crossref]

Fama, E.F., and K.R. French. 2015. A five-factor asset pricing model. Journal of Financial Economics 116: 1–22.
[Crossref]

Fama, E.F., and K.R. French. 2018. Choosing factors. Journal of Financial Economics 128: 234–252.
[Crossref]

Fama, E.F., and J.D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests. Journal of Political Economy 81: 607–636.
[Crossref]

Gibbons, M.R., S.A. Ross, and J. Shanken. 1989. A test of the efficiency of a given portfolio. Econometrica 57: 1121–1152.
[Crossref]

Harvey, C.R., Y. Liu, and H. Zhu. 2016. and the cross-section of expected returns. Review of Financial Studies 29: 5–68.
[Crossref]

Hou, K., C. Xue, and L. Zhang. 2015. Digesting anomalies: An investment approach. Review of Financial Studies 28: 650–705.
[Crossref]

Jones, P.N., and G.J. McLachlan. 1990. Algorithm AS 254: Maximum likelihood estimation from grouped and truncated data with finite normal mixture
models. Applied Statistics 39: 273–282.
[Crossref]

Liu, W. 2013. A New Asset Pricing Model Based on the Zero-Beta CAPM: Theory and Evidence. Doctoral dissertation, Texas A &M University.

Liu, W., J.W. Kolari, and J.Z. Huang. 2012. A new asset pricing model based on the zero-beta CAPM. Best paper award in investments at the annual
meetings of the Financial Management Association, Atlanta, GA, October.

Liu, W., J.W. Kolari, and J.Z. Huang. 2019. Creating superior investment portfolios. Working paper, Texas A &M University.

Kolari, J.W., W. Liu, and J.Z. Huang. 2021. A New Model of Capital Asset Prices: Theory and Evidence. Cham: Palgrave Macmillan.
[Crossref]
Kolari, J.W., J.Z. Huang, H.A. Butt, and H. Liao. 2022a. International tests of the ZCAPM asset pricing model. Journal of International Financial
Markets, Institutions &Money 79: 101607.

Kolari, J.W., W. Liu, J.Z. Huang, and H. Liao. 2022b. Further tests of the ZCAPM asset pricing model. Journal of Risk and Financial Management 15:
1–23.
[Crossref]

Linnainmaa, J., and M. Roberts. 2018. The history of the cross-section of stock returns. Review of Financial Studies 31: 2606–2649.
[Crossref]

Lintner, J. 1965. The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets. Review of Economics and
Statistics 47: 13–37.
[Crossref]

Markowitz, H.M. 1959. Portfolio selection: Efficient Diversification of Investments. New York, NY: Wiley.

McLachlan, G.J., and T. Krishnan. 2008. The EM Algorithm and Extensions, 2nd ed. New York, NY: Wiley.
[Crossref]

McLachlan, G. J., and D. Peel. 2000. Finite Mixture Models. New York, NY: Wiley Interscience.
[Crossref]

McLean, R.D., and J. Pontiff. 2016. Does academic publication destroy predictability? Journal of Finance 71: 5–32.
[Crossref]

Merton, R.C. 1973. An intertemporal capital asset pricing model. Econometrica 41: 867–887.
[Crossref]

Mossin, J. 1966. Equilibrium in a capital asset market. Econometrica 34: 768–783.

[Crossref]

Roll, R. 1977. A critique of the asset pricing theory’s tests, part I: On past and potential future testability of the theory. Journal of Financial Economics
4: 129–176.
[Crossref]

Ross, S.A. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13: 341–360.
[Crossref]

Sharpe, W.F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk. Journal of Finance 19: 425–442.

Stambaugh, R.F., and Y. Yuan. 2017. Mispricing factors. Review of Financial Studies 30: 1270–1315.
[Crossref]

Treynor, J.L. 1961. Market value, time, and risk. Unpublished manuscript.

Treynor, J.L. 1962. Toward a theory of market value of risky assets. Unpublished manuscript.

Footnotes
1 The material in the book is based on earlier academic studies of U.S. stock market returns by the authors in Liu et al. (2012, 2019) and Liu (2013) as
well as real-world investment services in collaboration with the Teachers Retirement System of Texas (TRS). Additionally, recent papers by Kolari et al.
(2022a, b) extend their previous work to a number of major industrialized countries as well as a longer U.S. sample period.

2 Interested readers can refer to the random matrix theory and Markowitz’s mean-variance portfolio proofs in their book.

3 We have simplified the notation somewhat here. The authors use $$f(\theta ) \sigma ^2_a$$ instead of simply $$\sigma ^2_a$$ in Eq. (10.2),
where $$f(\theta ) &gt; 0$$ is a complex function of other terms.

4 The following equilibrium conditions are implied by the ZCAPM:

1.
assuming all funds are invested in either $$I^*$$ or $$ZI^*$$, then $$\beta _{I^*,a} =\beta _{ZI^*,a} = 1$$ and $$Z^*_{I^*,a} = 1$$
or $$Z^*_{ZI^*,a} = -1$$, respectively;
2.
assuming no riskless asset, Eq. (10.8) reduces to Eq. (10.4) (i.e., $$\beta _{i,a}\equiv w_{I^*}+w_{ZI^*}=1$$); and
3.
assuming the restriction $$w_{f}&gt;0$$ (i.e., no borrowing at the riskless rate is allowed), then $$\beta _{i,a}&lt;1$$.
5 They are observable in the sense that they can be estimated with standard mean and variance statistical measures of market characteristics.

6 The authors estimated the ZCAPM to compute beta and zeta risk coefficients for thousands of U.S stocks (see Sect. 10.2 for an overview of the
empirical ZCAPM). In 1964 the ZCAPM was estimated for stocks which were then sorted into quintiles by their beta and zeta coefficients to form the 25
portfolios. Next, portfolios’ returns were computed in the out-of-sample month of January 1965. This process was rolled forward one month at a time to
generate monthly series of beta, zeta, and one-month-ahead returns for each portfolio from January 1965 to December 2018. For each portfolio, average
returns and the standard deviations of these monthly returns were computed. All portfolios were long-only with no short positions.

7 Originally invented by Dempster et al. (1977), EM regression is widely used in the sciences. See also studies by Jones and McLachlan
(1990), McLachlan and Peel (2000), and McLachlan and Krishnan (2008), among others. Wikipedia gives an excellent discussion of the EM algorithm
with literature citations.

8 See https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.

9 This portfolio is the smallest-size/lowest-BM (growth) stock portfolio with the lowest average realized excess return. It is well known that the three-
factor model has difficulty explaining this small, growth portfolio (e.g., see Fama and French (1996)).

10 We should mention that some researchers have found that many anomalous factors tend to disappear after their publication in finance journals (e.g.,
see McLean and Pontiff (2016), Linnainmaa and Roberts (2018), and others). Nonetheless, numerous anomalous factors persist over time.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
J. W. Kolari, S. Pynnönen, Investment Valuation and Asset Pricing
https://doi.org/10.1007/978-3-031-16784-3_11

11. Event Studies

James W. Kolari1 and Seppo Pynnönen2
(1) Mays Business School, Texas A&M University, College Station, TX,
USA
(2) Departent of Mathematics and Statistics, University of Vaasa, Vaasa,
Finland

James W. Kolari (Corresponding author)

Email: [email protected]

Seppo Pynnönen
Email: [email protected]

Supplementary Information
The online version contains supplementary material available at https://doi.
org/10.1007/978-3-031-16784-3_11

One of the most common applications of asset pricing models is event

studies.1 As discussed in Chapter 2, consistent with the Random Walk
Hypothesis (RWH), stock prices move randomly over time such that price
changes are independent of one another over time. This price behavior
supports the notion of efficient markets. To more rigorously test efficient
markets, Fama et al. (1969) introduced the concept of an event study by
examining the stock price reactions of companies to stock split
announcements. They wanted to know the speed of adjustment of stock
prices in response to news about stock splits by companies, which is an
indicator of market efficiency. Importantly, the authors developed a new
methodology to study the movement of stock prices over time. An event
study of stock splits was proposed using the market model form of the
CAPM to measure abnormal returns not explained by systematic market
risk. Abnormal returns are firm-specific or idiosyncratic returns after taking
into account systematic risk effects on returns. Empirical analyses of
residual error terms used to measure abnormal returns showed that stock
returns increase prior to stock splits and then level off thereafter. They
interpreted these findings to mean that stock splits signal to investors that
the firm will increase dividends in the future. Stock returns rise prior to
splits due to investors recognizing the increasing profitability of the firm.
The split itself signals that those expectations are confirmed by the firm.
Interestingly, firms that later decreased (increased) dividends experienced
lower (unchanged) stock returns after the split. These abnormal return
patterns confirmed that splits are dividend signals from the prospective
investors. Finally, stock prices reacted immediately on the day of split
announcements in line with market efficiency.
Over time researchers have conducted event studies to investigate the
stock market reaction to many different kinds of news announcements. For
example, many studies have examined firm-specific events besides stock
splits, including managerial decisions, earnings, and other accounting-
related announcements.2 Other studies look at economy-wide events,
including regulatory changes and macroeconomic announcements.3 Today,
event studies not only provide evidence on the question of market
efficiency but how investors perceive different kinds of information that
will impact stock prices.
This chapter reviews the traditional event study methodology
introduced by Fama, Fisher, Jensen, and Roll and developed by researchers
over more than 50 years. In this regard, event studies can be categorized as
short or long run. Typically, studies within a 1-year timeframe are
considered to be short run. Those extending 1-to-5 years after events are
long-run studies. As we will see, short- and long-run studies apply different
statistical methods. Also, long-run studies have become very controversial
as abnormal returns should not persist over many years in an efficient
market.

11.1 Short-Run Event Studies

Brown and Warner (1980, 1985) and MacKinlay (1997)4 have provided
excellent discussions and references with respect to traditional, short-run
event study analysis. In this section, we overview how asset pricing models
are used to conduct short-run event studies.

11.1.1 Basic Steps in an Event Study

Event study analyses rely upon asset pricing models. We next go through
the basic steps to conduct a short-run event study. We assume daily returns
but weekly or monthly returns can be used too.
1.
Define the event of interest—for example, an earnings announcement
by a company. Often the event must be operationalized more precisely.
For example, Campbell et al. (1997, Chapter 4) assigned each earnings
announcement to one of the three categories: good news, no news, and
bad news. If the actual earnings exceeded (fell below) expected or
forecasted earnings by 2.5% or more, the announcement was
designated as good (bad) news. Other earnings closer to expected were
considered no news.
2.
Collect a sample of firms that had the event. Relevant companies are
required to have sufficient data to estimate asset pricing model
parameters.
3.
Determine for each sample firm of the event day.
4.
Based on the event day, define the event window. The event date is
labeled as day 0 and other days around the event day comprise the
event window. Event days with negative values occur before the event
day, and positive values occur on post-event days. For example, days
mean one day before the event day, the event day, and the
next day after the event day.
The event window can be (for example) trading days around
the event day, i.e., consisting of 11 trading days , , , ,
, 0, 1, 2, 3, 4, 5. The event window can be longer if desired; for
example, Campbell et al. used an event window of trading days
around the announcement, or about one month before and one month
after the event (41 trading days).
5. Define the estimation window prior to the event window. With daily
data, the length of the estimation window may be (for example) 250
trading days or about one year Given the above 11 day event window
 trading days, or about one year. Given the above 11-day event window

from days to , trading days from to would

encompass the estimation window.
6.
Using an asset pricing model, compute the abnormal returns for each
firm on each event day in the event window. We discuss this
calculation in detail below.
7.
Make inferences about the behavior of the abnormal returns based on
appropriate statistical tools, including descriptive statistics, graphics,
and tests. Report the statistical findings and discuss their economic
significance also (i.e., the magnitudes of abnormal returns from the
perspective of investors).

11.1.2 Abnormal Returns

To evaluate the impact of an event on the stock returns of a company, we
need to compute the abnormal return (AR). The AR is defined as the
deviation of the actual return from the predicted return on each day in the
event window:
(11.1)
where is the actual return at time t, and is the predicted
return. The latter return is the conditional expected return given information
under normal conditions (i.e., without the event effect). Generally
speaking, is called the mean model that measures the normal
performance of returns. Summing up the abnormal returns over the event
window from to yields cumulative abnormal returns (CARs):

(11.2)

There are several ways to specify and estimate the mean model. The
simplest approach is to specify equal to the unconditional
expected return , which can be estimated by the sample mean
over the estimation window. A more common way to measure abnormal
returns is to use an asset pricing model. Following Fama et al. (1969), let’s
take the simplest case of the market model written as
, where is the proxy market return.5
Extending the market model approach, event studies have utilized a
multifactor model written as ,
where the information set is based on multifactor values (e.g., the
Fama and French [1993, 2015] three- and five-factor models with
). Interestingly, Brown and Warner (1985) have found that
statistical results in event studies are not critically dependent on the chosen
asset pricing model. Hence, short-run event study results are not very
sensitive to the asset pricing model employed.
Given the factor model, the regression parameters and ,
are estimated by ordinary least squares (OLS) using the
estimation window returns for each sample stock . Estimates
for the abnormal returns in the event window are prediction errors:
(11.3)
where t is the event time within the event window for
different assets.
In short-run event studies, statistical tests of event effects can be broadly
divided into two categories: parametric and nonparametric. Parametric
tests rely on the assumption that returns follow some parametric probability
distribution like the Gaussian normal distribution. Nonparametric tests relax
this assumption and therefore tend to be more robust. Even so, this
robustness does not come without some cost. For instance, if the underlying
assumptions of the parametric return generating process hold, parametric
tests tend to be more powerful, which means that they will detect the event
effect with higher probability than nonparametric test if indeed there is a
significant event (i.e., Type II error is reduced). Conversely, if the
underlying assumptions are violated, parametric tests can be prone to false
alarms in the sense that they detect significant event effects when there are
none (i.e., Type I error is increased).6 Because the vast majority of event
studies utilize parametric testing, we focus on this approach here. For
interested readers, Appendix A provides a short introduction to
nonparametric approaches.

11.1.3 Parametric Testing

Statistical testing for an event effect involves testing whether the population
mean of the abnormal returns deviates from zero after a news
announcement. That is, we want to know whether the observed deviation of
actual returns from predicted returns is due to natural random variation in
stock returns or due to new information that causes stock prices to change.
The null hypothesis is:
(11.4)
where is the population mean of abnormal returns across all event firms.
Similarly, denoting the population mean of cumulative abnormal returns as
over event days from to as , the null
hypothesis of no cumulative event effect is:
(11.5)

Non-Standardized Return Tests

A generic form of the t-statistic for testing the null hypotheses in
Eqs. (11.4) and (11.5) is the traditional t-ratio:

(11.6)

where is the sample mean of ARs (or CARs) of sample event firms, and
is the standard error of the sample mean.
Assuming that the event windows of different firms are not overlapping
in calendar time, the t-statistic in Eq. (11.6) for testing the null hypothesis
(11.4) for ARs becomes:

(11.7)

where

(11.8)

is the sample average of abnormal returns on event day , and

is the standard error of . Similarly, for testing the null
hypothesis (11.5) for CARs, the t-ratio in Eq. (11.6) becomes:
 (11.9)

where

(11.10)

and is the standard error.7 If the event windows are

completely nonoverlapping, the abnormal returns between firms can be
considered independent. In this case, the central limit theorem (CLT)
implies that the null distributions of both of these statistics are
approximately standard normal, or N(0, 1).
There is considerable literature on the development of robust t-statistics
in event studies. These studies seek to provide more accurate tests of event
effects that measure whether stock price changes are significant or not. For
example, Harrington and Shrider (2007) noticed that news events can cause
the variances of stock returns to increase. The authors recommended an
adjustment to standard errors in the t-statistic to account for this issue.
Another potential issue that can affect standard errors is returned
autocorrelation (i.e., correlation of stock returns over time). Because the
above t-ratios utilize returns in the event window and not the estimation
window, they are robust to both event-induced variance and
autocorrelation.8

Standardized Return Tests

It was soon observed that material statistical power gains could be achieved
by using standardized abnormal returns rather than original returns. The
best known statistical tests are the Patell (1976) z-statistic and Boehmer
et al. (1991) t-statistic referred to as the BMP-test. In general, these more
sophisticated tests better measure the statistical significance of stock returns
in response to news events. While somewhat complex, these tests can be
fairly easily computed by means of statistical software programs.
Standardized abnormal returns are defined as:

(11.13)
where
(11.14)
is the sample standard deviation of in which is the OLS residual
standard deviation of the estimation window residuals and is the
prediction error correction to account for sampling errors in the OLS
regression coefficient estimates. Standardized cumulative abnormal returns
(SCARs) over event days are defined in the same manner:

(11.15)

where
(11.16)
in which is the same as in Eq. (11.14).9
For SARs computed using the market model, Patell (1976) proposed the
following test statistic:

(11.17)

where is the length of the estimation window of the ith stock. By the
CLT in large samples, we know that .
The Patell statistic does not account for possible event-induced
variance. To fix this problem, Boehmer et al. (1991) proposed to utilize the
event time t estimated standard error in the test statistic. As such, their test
statistic is analogous to the t-ratios in Eqs. (11.7) and (11.9) with s and
s replaced by their standardized counterparts. Hence, for s, we
now have:

(11.18)

where is the standard error of computed similarly to

Eq. (11.11) using s. Again, assuming independence, the CLT implies
that the null distribution of the test statistic is approximately N(0, 1). Due to
its robustness to event-induced variance, the BMP statistic has become one
of the most popular test statistics in event studies.10
If the events are clustered in calendar time, meaning that all or a large
part of the event firms share the same event day in calendar time, the above
statistics tend to overreject the null hypothesis when there is no event effect.
This bias is due to the contemporaneous correlation of stock returns. The
use of appropriate mean models to estimate abnormal returns was
previously thought to alleviate this symptom. However, Kolari and
Pynnönen (2010) demonstrated that even a low cross-sectional correlation
causes incremental overrejection as the sample size grows. Consequently,
they proposed the following adjustment to the BMP statistic (adjusted
BMP) when all firms share the same event day in calendar time:

(11.19)

where is an estimate of the average cross-sectional correlation of the

abnormal returns. This test can be compted using either the estimation
period or event period or these combined periods. If the event days are
partially clustered, meaning that groups of firms share some but not all
event days in calendar time, Kolari et al. (2018) proposed the following
adjustment:

(11.20)

where is the fraction of overlapping calendar days in the computed

abnormal return. If all event days are different in calendar time, and
reduces to . However, if all event days are the same, and
reduces approximately to 11
. Altogether, these test statistics
provide robust statistical tools for researchers and professionals to conduct
short-run event studies.

11.1.4 Examples
Example 1: To demonstrate event study testing of abnormal returns, we
employ daily stock returns for a sample of 50 seasoned equity offerings
(SEOs) by U.S companies. We collect this sample of events in the period
1985– 2015. SEOs raise equity capital for firms to deploy in various capital
expenditures (e.g., merger and acquisition bids, fixed plant expenses, etc.).
They are highly visible events as new information about the future activities
of a firm is revealed to investors. The event period (or event window) is
days around the event day 0. The estimation period is 250 days prior
to the event period, i.e., days from to .
Table 11.1 reports results using the test statistics for event day 0 as well
as the cumulative abnormal return windows , , ,
and . In this sample the event days are not clustered (i.e., SEOs
for different firms happen on different days), so no cross-sectional
correlation adjustments are needed. The event day return is %, which
is economically large and statistically significant in terms of the non-
standardized test as well as standardized tests of the Patell and BMP.
None of the other event windows shows significant abnormal returns. We
infer that these results support market efficiency wherein market
information is rapidly absorbed in stock prices and returns without delays or
pre-event leakages. In general, investors viewed SEOs negatively as stock
returns materially decreased on the event day.
Table 11.1 Short-run event study results for 50 SEO events of U.S. firms from 1985 to 2015

Event window
Event day
AR/CAR(%) −2.54 −2.31 −1.21 0.27 −3.84
$$t_{ CAR }$$ −1.95 −1.65 −0.98 0.20 −1.52
p-value 0.056 0.105 0.331 0.840 0.134
$$z_{\textrm{patell}}$$ −4.50 −2.31 −1.54 1.83 −1.09
p-value 0.000 0.021 0.125 0.067 0.274
$$t_{\textrm{bmp}}$$ −2.20 −1.40 −1.41 1.33 −1.12
p-value 0.033 0.196 0.165 0.190 0.269

Example 2: We next demonstrate event study testing for economy-wide

announcements. Instead of focusing on individual firms, we investigate the
impacts of announcements on industries. Using standard regression
software, we consider selected announcements related to financial crises
that occurred in 2008 and 2020.12 Our investigation focuses on how news
events impacted U.S. stock returns for different industry portfolios. We
selected the following major news announcements in 2008 and 2020:
2008 financial crisis news events
January 11: Bank of America agreed to buy failing savings bank
Countrywide Financial for about $4 billion.
March 16: The Federal Reserve agreed to guarantee $30 billion in assets
held by securities firm Bear Stearns in connection with its government-
sponsored sale to JP Morgan Chase.
September 15: Lehman Brothers, a large securities firm, is not bailed out
by the government and therefore files for bankruptcy. Bank of America
agreed to purchase failing securities firm Merrill Lynch for $50 billion.
October 3: Congress passed a bailout program for big banks (known as
TARP). Wells Fargo agreed to buy failing Wachovia for about $14.8
billion.
2020 COVID-19 pandemic news events
March 11: The World Health Organization (WHO) declared COVID-19 a
global pandemic.
March 26: The U.S. Congress passed the Coronavirus Aid, Relief, and
Economic Security (CARES) Act. This legislation provided $2 trillion in
aid to hospitals, small businesses, and state and local governments.
We downloaded daily stock returns from Kenneth French’s online data
library13 for four industries: consumer discretionary, manufacturing,
communications, and health. Days in the analyses are denoted as follows:
$$\tau _1$$ = start estimation window, $$\tau _2$$ = end
estimation window, $$\tau _3$$ = event day, and $$\tau _4$$ =
end event window. As such, the event window consists of days
$$\tau _2 + 1, \ldots , \tau _4$$ and the event day is a day within the
event window, i.e., $$\tau _2 &lt; \tau _3 \le \tau _4$$.
To run the event study analyses, special event-study software is not
needed. We can easily adapt standard regression/econometric software to do
the analyses. We define a dummy variable for each day within the event
window and add the dummies to the regression model. Here we use the
Fama-French five-factor model as the mean model (normal return) for
industry p and enhance the model with the event window dummy variables
as follows:
 $$\begin{aligned} \begin{aligned} R_{pt}^e ={}&amp;\alpha _p (11.21)
+ \beta _pR_m^e + s_p SMB _t + h_p HML _t + r_p RMW + c_p
CMA _t + \delta _{\tau _2 + 1}D_{\tau _2 + 1} \\&amp;+ \cdots +
\delta _{\tau _4}D_{\tau _4} + \epsilon _{pt}, \end{aligned}
\end{aligned}$$
where $$R_p^e = R_p - R_f$$ is the excess return for an industry
portfolio over the riskless Treasury bill rate $$R_f$$,
$$R_m^e = R_m - R_f$$ is the excess return for the market proxy,
and $$SMB$$, $$HML$$, $$RMW$$, and
$$CMA$$ are the well-known Fama-French zero-investment factor
portfolios. Dummy variables $$D_{\tau _2 + j}$$ equal 1 on event
days $$\tau _2 + j = t$$ and 0 otherwise, which correspond to the
days in the event window. Importantly, the respective coefficients
$$\delta _{\tau _2 + j}$$ measure event effects on each day within the
event window.
Using standard regression programs available in SAS, Stata, R, or other
statistical software packages, regression (11.21) can be estimated. The
$$\delta$$-estimates equal the prediction errors of the regression
model and therefore comprise the needed abnormal returns in Eq. (11.3) to
measure the event effects. For instance, testing for the event day abnormal
return reduces to testing whether the corresponding coefficient (or
$$\delta _{\tau _3}$$) is zero. Testing cumulative abnormal returns,
such as CAR (+1, +5), tests whether the sum of the respective coefficients is
zero. Given $$\tau _3$$ denotes the event day, the null hypothesis is:
$$\begin{aligned} H_0: \delta _{\tau _3 + 1} + \cdots + \delta
(11.22)
_{\tau _3 + 5} = 0. \end{aligned}$$
Using the above regression techniques, Table 11.2 reports the event
study results. Again, the estimation window is 250 trading days before the
event window of $$\pm 10$$ trading days around the event day.
Panel A shows cumulative abnormal returns $$\textit{CAR}(-5, -1)$$
for five days before the event day, Panel B tests abnormal returns on event
day 0, and Panel C tests $$\textit{CAR}(+1, +5)$$ for five days after
the event day. The results show that event effects vary across industries. For
instance, in Panel C for the October 3, 2008 announcement of Wells Fargo’s
purchase of Wachovia, there were negative $$\textit{CAR}(+1, +5)$$
s in the manufacturing and health industries but no effects on the
communications industry. WHO’s COVID-19 declaration on March 11,
2020 appears to have affected all industries but in different ways. The
consumer and manufacturing industries had highly significant negative
$$\textit{CAR}(+1, +5)$$s, but the communications and health
industries had significant positive $$\textit{CAR}(+1, +5)$$s.
Table 11.2 Abnormal returns for four industries using the Fama and French five-factor model:
Selected news announcement dates during the 2008 financial crisis and 2020 COVID-19 crisis

Event date (Financial crisis) Event date (COVID-19)

Industry Jan 11, Mar 16, Sept 15, Oct 3, Mar 11, Mar 26,
2008 2008 2008 2008 2020 2020
Panel A: CAR (−5, −1)
Consumer 1.79 −0.36 1.08 1.28 1.03 0.40
(0.002)*** (0.604) (0.219) (0.160) (0.124) (0.573)
Manufacturing −1.15 0.66 −1.20 −3.54 −2.56 0.60
(0.146) (0.426) (0.281) (0.003)*** (0.000)*** (0.433)
Communications −2.97 0.36 1.48 0.09 0.01 −1.52
(0.000)*** (0.639) (0.134) (0.929) (0.981) (0.010)***
Health 4.35 −0.77 1.00 1.90 0.07 −5.02
(0.000)*** (0.418) (0.395) (0.112) (0.942) (0.000)***
Panel B: Event day 0 abnormal returns
Consumer −0.99 −0.01 0.97 −0.63 −0.58 −0.36
(0.000)*** (0.979) (0.017)** (0.119) (0.040) (0.231)
Manufacturing 0.01 0.07 −1.23 1.02 −1.31 1.18
(0.977) (0.843) (0.017)** (0.055)* (0.000)*** (0.000)***
Communications −0.47 0.73 0.25 −0.31 0.74 0.02
(0.107) (0.036)** (0.582) (0.491) (0.001)*** (0.948)
Health 1.34 0.07 −0.16 −0.54 −0.14 1.06
(0.000)*** (0.868) (0.772) (0.305) (0.736) (0.012)**
Panel C: CAR (+1, +5)
Consumer 1.33 2.03 −3.16 −0.81 −1.97 0.16
(0.019)** (0.005)*** (0.001)*** (0.408) (0.004)*** (0.803)
Manufacturing −1.30 −3.62 4.24 −6.47 −3.27 1.39
(0.099)* (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.053)*
 Event date (Financial crisis) Event date (COVID-19)
Industry Jan 11, Mar 16, Sept 15, Oct 3, Mar 11, Mar 26,
2008 2008 2008 2008 2020 2020
Communications 1.05 −0.32 −2.28 1.85 3.01 0.02
(0.109) (0.680) (0.027)*** (0.090)* (0.000)*** (0.965)
Health −1.17 0.21 −1.59 −6.93 3.32 1.94
(0.102) (0.830) (0.195) (0.000)*** (0.001)*** (0.037)**

Abnormal returns are computed using the Fama and French five-factor
model:
$$R_p^e = \alpha _p + \beta _pR_m^e + s_p SMB + h_p HML + r_p
RMW + c_p CMA + \epsilon _p,$$
where $$R_p^e = R_p - R_f$$ is the excess return over the riskless
Treasury bill rate for an industry portfolio, $$R_m^e = R_m - R_f$$ is
the excess return for the market proxy, and $$SMB$$, $$HML$$,
$$RMW$$, and $$CMA$$ are the zero-investment portfolio factors.
The estimation window is 255 trading days before the beginning of the
event window containing $$\pm 10$$ trading days around event day 0
Note p-values for the level of significance are shown in parentheses: * =
10*, ** = 5%, and *** = 1%
Data source https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_
library.html

Fig. 11.1 Cumulative abnormal returns for selected industries in response to 2008 financial
crisis news events

Fig. 11.2 Cumulative abnormal returns for selected industries in response to 2020 COVID-19 crisis
news events

Figures 11.1 and 11.2 graph the industry CARs in the 21-day event
windows for the news events. Notice in particular the larger negative effects
of the financial crisis announcement on October 3, 2008 (Fig. 11.1, Panel
D) and COVID-19 announcement on March 11, 2020 (Fig. 11.2, Panel A)
with respect to the manufacturing sector compared to the other sectors.
Interestingly, this industry seems to anticipate these events by a few days.
Thus, investors were expecting these events to some extent, which is
consistent with an efficient market.

11.2 Long-Run Event Studies

Both short- and long-run event studies are useful in terms of testing market
efficiency. As we have seen, short-run studies focus on immediate effects
and price adjustment in the days surrounding the event. By contrast, long-
run event studies seek to capture potential persistence in event effects over
time horizons from months to several years. Of course, persistent return
effects over time are inconsistent with market efficiency. In an efficient
market, investors should adjust returns to risks such that no abnormal
returns (or differences between actual and predicted returns) exist in the
long run.
While short-run event study methodology is well established, long-run
event study methods continue to develop. Major long-run event study issues
include proper normal expected return definitions and appropriate risk
adjustment. It is reasonable to believe that a major firm event (e.g., a
seasoned equity offer) will change firm fundamentals and therefore its risk
profile. Also, in contrast to the short-run event studies, the test statistic
specification in long-run studies is highly sensitive to the benchmark asset
pricing model to measure predicted returns. The simple reason is that,
within only a few days in short-run studies, a small bias in the predicted
return is indiscernible from random noise. However, these small
measurement biases can accumulate as the time horizon grows to several
years, which can cause both statistically and economically large differences.
Finally, risk shifts are likely due to major corporate events, which adds to
potential measurement problems in estimating predicted returns. For these
reasons, it is possible that an apparent long-run abnormal return is due
simply to measurement problems.
Current long-run event studies utilize two different methods to assess
post-event (risk-adjusted) abnormal return behavior: (1) buy-and-hold
abnormal returns, (2) and calendar time abnormal returns. We next review
these two test approaches.

11.2.1 Buy-and-Hold Abnormal Return Approach

Ikenberry et al. (1995), Barber and Lyon (1997), and Lyon et al. (1999)
introduced the buy-and-hold abnormal return (BHAR) approach. This
approach finances the investment in event firms by selling short similar
non-event firms and holding the position to the end of the period (e.g., five
years). A similar or matching non-event firm is called the control firm.
Given the matching control firm, a T-month BHAR is defined as
$$\begin{aligned} BHAR _{iT} = \prod _{t = 1}^T (1 + R_{it}) -
(11.23)
\prod _{t = 1}^T(1 + R_{it}^c), \end{aligned}$$
where $$R_{it}$$ is the return on the event firm i in post-event month t,
and $$R_{it}^c$$ is the return on the matching non-event firm. Often the
control firm is defined as a firm with similar size and book to market ratio
characteristics as the event firm.
Under the null hypothesis of no event effect, the event firm and its
match are expected to have similar expected returns, such the expected
value of $$BHAR _{iT}$$ is zero. Given a sample of n event firms,
the t-ratio for testing the null hypothesis is
$$\begin{aligned} t_{ BHAR } = \frac{\overline{ BHAR }}
(11.24)
{\mathrm {s.e}(\overline{ BHAR })}, \end{aligned}$$
where $$\overline{ BHAR }$$ is the sample mean, and
$$\mathrm {s.e}(\overline{ BHAR })$$ is the standard error.14
Unfortunately, this test statistic has two Achilles’ heels. One possible
problem is the independence assumption, which requires that the event
windows of all firms are completely non-overlapping in calendar time. In a
long window extending (for example) 5 years, this assumption may well not
hold. For this reason, because stock returns tend to be positively correlated,
the standard error in t-ratio (11.24) becomes downward biased causing
overrejection of the null hypothesis. A second problem is that BHARs are
prone to incomplete risk adjustment. Some adjustment for risk is
accomplished by subtracting control firm returns from event firm returns.
But the two firms can plausibly differ in terms of a number of important
market risks. Recognizing this drawback, Fama (1998) has advocated for
risk adjustment in computing abnormal returns, which favors the calendar
time abnormal return (CTAR) approach to be discussed next. Despite its
shortcomings, the BHAR approach has gained popularity due to its focus on
investor experience. That is, BHAR is based on compounded returns earned
by investors in a holding period.

11.2.2 Calendar Time Abnormal Return Approach

Introduced by Jaffe (1974), calendar timeabnormal returns (CTARs) are
risk-adjusted returns defined in terms of a calendar time portfolio. For
example, for a 5-year investment horizon, the month t return of an equal-
weighted calendar time portfolio is the average of returns of those stocks
that have had the event within the 5-year period. Applying the Fama-French
three-factor model to adjust for market, size, and value risk factors, we
have15:
$$\begin{aligned} R_{pt} - R_{ft} = \alpha _p + \beta _p(R_{mt}
(11.26)
- R_{ft}) + s_p SMB _t + h_p HML _t + e_{pt}, \end{aligned}$$
where $$R_{pt}$$ is the calendar time portfolio return, $$R_{ft}$$ is
the riskless rate, $$R_{mt}$$ is the proxy market return, $$SMB$$ is
the small minus big factor, $$HML$$ is the high minus low value factor,
and $$e_{pt}$$ is the error term. Notably, the intercept $$\alpha _p$$
is the calendar time abnormal return (CTAR). In an efficient market in
which all risks are impounded in asset prices, no abnormal return should
persist in the long run. Hence, if $$\alpha _p$$ is nonzero, we can infer
that an abnormal returns exist that is not explained by market risks.
It should be mentioned that a major controversy in finance exists with
respect to long-run abnormal returns. As already mentioned, assuming
efficient markets, all abnormal returns should be zero in the long run after
proper risk adjustment. Hence, the presence of long-run abnormal returns
suggests that market risks are not completely adjusted in computing
predicted turns from an asset pricing model. On the other hand, assuming
that inefficiency can exist in markets, behavioralists have argued that long-
run abnormal returns reflect over- and underreactions by investors to market
information due to psychological reasons. What do you think? What
explains long-run abnormal returns?
It is important to note that Boehme and Sorescu (2002) adapted the
CTAR approach to control firms in a manner similar to the BHAR approach.
More specifically, they subtracted matched control firm returns from event
firm returns on the left-hand-side of Eq. (11.26). By doing so, a nonzero
$$\alpha _p$$ could be interpreted as an abnormal return, as opposed
to mispricing error in the asset pricing model. Since mispricing error would
be present in both event and control stock returns, they will cancel out in
the event minus control stock return.
The null hypothesis of no (long-run) abnormal return reduces to testing
whether $$\alpha _p$$ differs from zero:
$$\begin{aligned} H_0 : \alpha _p = 0. \end{aligned}$$ (11.27)
The t-ratio is:
$$\begin{aligned} t_{\hat{\alpha }} = \frac{\hat{\alpha }_p}
(11.28)
{\mathrm {s.e}(\hat{\alpha }_p)}, \end{aligned}$$
which tests for the significance of abnormal returns in the 5-year event
period. A significant $$\alpha _p$$ estimate fails to accept (or rejects)
the null hypothesis of no event effect.
Unfortunately, BHAR and CTAR often produce different inferences. As
discussed above, BHAR emphasizes investor holding period returns,
whereas CTAR is a risk-adjusted abnormal return measure based on market
efficiency. Since long-run studies typically span many months, they can be
used to investigate risk shifts of the sample firms in response to major news
events. It is possible that both returns and risks change due to the impacts of
events on firms. The following real-world example for testing abnormal
returns illustrates these test approaches.

11.2.3 Example
Here we compare the BHAR and CTAR approaches to long-run event
studies based on a sample of 200 seasoned equity offerings (SEOs) among
U.S. firms from 1985 to 2015. Using a 5-year post-event period, the total
sample period extends to the end of 2020. Comprehensive analyses of this
event and other major corporate events are reported by Kolari et al. (2021).
For BHAR, control firms are selected by matching size (market
capitalization) and book-to-market (BM) ratio characteristics to event firms
using CRSP and Compustat databases. Firm size is calculated at the end of
December prior to the SEO announcement. The BM ratio is calculated at
the end of the year prior to the event.
Table 11.3 reports the results for the event month as well as post-event
months (i.e., 6-month as well as 1-, 2-, 3-, and 5-year periods). BHAR
abnormal returns are the excess returns over respective holding periods, and
CTAR abnormal returns are average returns per month. The signs of
abnormal returns computed using these two methods are the same for the
most part. One exception is 6 months, in which BHARs are approximately
4% but CTARs equal −0.31% monthly return or about −1.8% over six
months. However, both of these abnormal returns are not statistically
significant. A major difference in results occurs in the 1-year post-event
period, wherein BHARs equal an insignificant −4.77% but CTARs equal a
significant −0.81% per month or about −9.7% per year. Looking at other
results, both approaches lead to the same inferences of significant 2-year
abnormal returns and insignificant 3- and 5-year abnormal returns. In the
post-event 5-year period, BHARs and CTARs suggest that SEO firms
underperform their matches by an economically meaningful −22.1
percentage points and −0.28 percentage points per month or about −16.8%
over 5 years, respectively. While economically large in magnitude, as
observed by Kothari and Warner (2006), both approaches suffer from weak
power that causes problems in detecting statistically significant events. In
general, our results suggest that long-run abnormal returns occurred in
response to SEOs by firms. Is the cause behavioral or is there a risk
explanation for these negative abnormal returns? Alternatively, do
managers sacrifice shareholder wealth to achieve their goals for the firm?
Table 11.3 Long-run event study results for 200 SEO events of U.S. firms from 1985 to 2015

SEO Post-event months/years

Month 6 months 1 year 2 years 3 years 5 years
BHAR(%) −2.06 3.98 −4.77 −17.03** −13.28 −22.10
t-value −1.09 1.01 −0.94 −2.15 −1.55 −1.53
CTAR(%) −0.12 −0.15 −0.81** −0.75** −0.37 −0.28
t-value −0.08 −0.31 −2.11 −2.40 −1.35 −1.15
Obs (BHAR) 200 200 200 200 200 200
Obs (CTAR) 145 337 367 369 369 369
N months 200 1,199 2,393 4,584 6,334 8,767

Lastly, the number of observations used to compute CTARs can be used

to approximate the fraction of overlapping event months. From the SEO
month column in Table 11.3, we see that 55 companies share the same event
month.
11.3 Summary
This chapter reviewed basic steps for conducting short- and long-run event
studies. Important steps include specification of the nature of the event,
sample firms, event dates, event window, estimation window, and asset
pricing model. Based on the empirical results, inferences are made
regarding the statistical and economic significance of abnormal returns.
Abnormal returns are the difference between actual and predicted returns.
An asset pricing model is employed to estimate predicted returns. Short-run
event studies typically focus on high frequency returns, such as daily
returns. Parametric t-tests for the significance of abnormal returns use either
non-standardized or standardized returns. Popular short-run tests of
standardized abnormal returns have been developed by Patell (1976),
Boehmer et al. (1991), and Kolari and Pynnönen (2010).
We demonstrated short-run event study tests using seasoned equity
offerings (SEOs) announcements of firms within an 11-day event window.
No significant abnormal returns were detected, which suggested efficient
market reactions to SEO news announcements. We also investigated
economy-wide announcements related to the 2008 financial crisis and 2020
COVID-19 pandemic. Among four different industries, we found that these
news announcements had mixed positive and negative cumulative abnormal
return (CAR) effects depending on the industry. Also, it appeared that
investors anticipated some news announcements consistent with an efficient
market. Graphical analyses of news events were shown to be helpful in
visualizing the results of short-run event studies.
Long-run studies utilize lower frequency monthly data that can extend
for years after events. We covered two different methods for assessing post-
event abnormal returns: (1) the buy-and-hold abnormal return
(BHAR) approach by Ikenberry et al. (1995), Barber and Lyon (1997), and
Lyon et al. (1999), and the calendar time abnormal return (CTAR) approach
by Jaffe (1974). While short-run methods are well established in the
literature, long-run methods continue to develop. BHARs capture long-run
abnormal returns in a holding period from the perspective of an investor,
whereas CTARs are risk-adjusted measures based on the intercept (
$$\alpha$$) term in an asset pricing model. We demonstrated these
approaches by means of seasoned equity offering (SEO) announcements.
The analyses showed that different results can be produced by these
alternative test methods. In general, we found negative long-run abnormal
returns after SEOS, which could be attributable to behavioral or risk
explanations. Since long-run studies can span many months, they can be
used to investigate risk shifts of the sample firms in connection with major
news events also.

Questions
1.
Explain the concepts of event date, event time, event window, and
estimation window.
2.
Explain normal return, abnormal return, and cumulative abnormal
return.
3.
Discuss pros and cons of standardized event study tests with relation to
non-standardized tests.
4.
Suppose that you want to test the null hypothesis of no mean effect
when event days are completely clustered. The sample size is
$$n = 100$$ firms and the test statistic applied is BMP due to its
good sample properties. You estimate the value of
$$t_{\textrm{bmp}} = 2.7$$ with $$p = 0.007$$ in two-sided
testing. As a result, you infer that the event effect is highly statistically
significant. Since the average cross-sectional correlation of the
abnormal returns is $$\bar{r} = 0.02$$, you argue that the small
average cross-sectional correlation will not materially change the test
results. Is this right?
5.
In the text it was noted that the BHAR approach in long-run event
studies is vulnerable to cross-sectional correlation. Is it vulnerable to
autcorrelation or heteroskedasticity of the returns too?
6.
Consider a long-run event study of SEO events with abnormal returns
defined as the difference between the returns of the event firm and its
matched control firm. What would you consider to be a major problem
with this definition of the abnormal return?
7. Discuss some of the major pros and cons of nonparametric rank tests in
 event studies (see Appendix A).

Problems
1.
A key requirement for efficient capital markets is that prices fully and
instantaneously reflect all available relevant information. Consider the
event of an earnings announcement. Suppose that the current price of a
stock is 100. Sketch a stock price process for $$\pm 5$$ days around
the event (using end of day prices) to reflect:
a.
an efficient capital market when the announcement is

i. a positive surprise with +5% price effect

ii. a negative surprise with −5% price effect
iii. neutral (i.e., according to analyst expectations) with no price
effect

b.
leakage of information before the event day when the surprise is

i. positive (+5%)
ii. negative (−5%)

c.
a price process as in (a) when the full price adjustment takes place
over a few days after the event.
2. The table below reports event day abnormal returns, standard
deviations of the estimation window abnormal returns (i.e., OLS
residuals of the market model), and rank number of the event day
abnormal return for each stock relative to its estimation window
abnormal returns in a sample of $$n = 30$$ stocks. The estimation
window is $$T_e = 200$$ days so that $$T = T_e + 1$$ (i.e.,
estimation window returns $$+$$ event return) is the total number of
returns for each stock. The smallest return for each stock assumes rank
number 1 and the largest 201. Using this ranking approach, the rank
 number 164 for the first stock indicates that the event day abnormal
return of 1.21% is the 164th largest among the 201 returns from the
combined estimation period and the event day. Test the null hypothesis
of zero abnormal return using each of the following statistics:

a. $$t_{ AR }$$
b. $$z_{\textrm{patell}}$$
c. $$t_{\textrm{bmp}}$$
d. $${}^*$$ $$z_{\textrm{sgn}}$$
e. $${}^*$$ $$t_{\textrm{grank}}$$.

$${}^*$$Optional: In the optional problem (d), assume that the

$$d_{it} = 1 / T_e$$ in Eq. (A11.1). In optional problem (e), when
computing Eq. (A11.4), you can use the theoretical standard error that in
this case is:
$$\begin{aligned} s_{\bar{U}} = \sqrt{\frac{1}{nT}\sum _{i =
1}^T(i/(T + 1) - 1/2)^2}, \end{aligned}$$
where $$T = T_e + 1 = 201$$. Using this equation, the result is
$$s_{\bar{U}} = 0.0524$$ (rounded to four decimal places).

Stock AR Std dev Rank

1 1.21 1.04 165
2 0.27 1.01 134
3 1.01 1.13 162
4 −0.81 1.03 47
5 0.32 0.84 116
6 −0.87 0.98 60
7 −0.41 1.20 82
8 1.13 0.94 159
9 1.16 1.17 161
10 −0.20 0.98 84
11 0.89 1.17 153
Stock AR Std dev Rank
12 0.97 1.08 150
13 1.68 1.14 180
14 1.13 0.99 171
15 1.79 1.05 186
16 0.42 1.18 111
17 −1.64 1.10 19
18 −0.71 0.96 56
19 0.91 1.20 154
20 0.29 1.15 124
21 2.10 1.18 187
22 0.22 1.06 114
23 0.19 1.03 116
24 1.32 1.06 172
25 2.29 0.92 199
26 −3.05 0.95 1
27 0.77 0.85 161
28 0.86 0.97 162
29 −1.38 0.97 14
30 1.32 0.99 175

Appendix A: Nonparametric Testing

Section 11.1.3 discussed parametric tests wherein data are assumed to be
generated from a particular process like the normal distribution. If this
assumption holds, the implied test statistics are optimal for statistical
testing. Even if the normality of returns does not hold, under fairly general
assumptions, the Central Limit Theorem (CLT) guarantees incrementally
accurate test results as the sample size grows. Nonparametric approaches
are free from any assumption about the specific distribution of the returns.
In this respect, the sign test and the ranks test are the two major categories
of nonparametric tests.
The sign test assumes that the (cumulative) abnormal returns are cross-
sectionally independent and, under the null hypothesis of no event effect,
the positive and negative sign of the CAR is equally likely. That is,
$$P( CAR \le 0) = P( CAR &gt; 0) = 1/2$$. Defining dummy
variables $$D_i = 0$$ when $$CAR_i &lt; 0$$ and
$$D_i = 1$$ when $$CAR_i \ge 0$$, and denoting
$$\bar{D} = \sum _{i = 1}^nD_i / n$$, then under the null hypothesis
of no event effect $$E(\bar{D}) = 1/2$$ and
$$\textrm{var}(\bar{D}) = 1/(4n)$$. Subsequently, we can define a
test statistic
$$\begin{aligned} z_{\textrm{sgn}} = 2 \sqrt{n} (\bar{D} - 1/2),
(A11.1)
\end{aligned}$$
which under the null hypothesis of no event effect is asymptotically N(0, 1)
distributed by the CLT.
The basic assumption of the sign test is that the return distribution is
symmetric. For this reason, it is vulnerable to skewness of the distribution,
in which case under the null distribution the probabilities of negative and
positive signs deviate from one half.
Corrado (1989) developed a test based on ranks of returns. Kolari and
Pynnönen (2011) refined the approach to cover both single day and
cumulative abnormal returns. They defined estimation window standardized
returns as $$SAR _{it} = AR _{it} / s_i$$, in which
$$AR _{it}$$s are estimation window OLS residuals, and
$$s_i$$ is the respective OLS standard error. For an event window,
they defined standardized cumulative abnormal returns from event day
$$\tau _1$$ to $$\tau _2$$ as
$$\begin{aligned} SCAR _{i}(\tau _1, \tau _2) = \frac{ CAR _i(\tau _1,
\tau _2)}{s_i(\tau _1, \tau _2)}, \end{aligned}$$
where $$s_i(\tau _1, \tau _2)$$ is the prediction error corrected OLS
standard deviation of the cumulative abnormal return (see Kolari et al.
2018). Re-standardizing $$SCAR _i(\tau _1, \tau _2)$$ by their cross-
sectional standard deviation helps to account for possible event-induced
variance. The re-standardized SCARs are defined as
$$\begin{aligned} SCAR _i^*(\tau _1, \tau _2) = \frac{ SCAR _i(\tau _1,
\tau _2)}{s(\tau _1, \tau _2)}, \end{aligned}$$
where
$$\begin{aligned} s(\tau _1, \tau _2) = \sqrt{\frac{1}{n-1}\sum _{i =
1}^n( SCAR _i(\tau _1, \tau _2) - \overline{ SCAR }_{\tau _1, \tau
_2})^2}. \end{aligned}$$
Using these together with the estimation window abnormal returns, Kolari
and Pynnonen defined the following generalized standardized abnormal
returns (GSARs):
$$\begin{aligned} GSAR _{it} = \left\{ \begin{array}{ll} SAR
_{it}, &amp;{} \text {for the estimation window returns}\\ SCAR
(A11.2)
_i^*(\tau _1, \tau _2), &amp;{} \text {for the cumulative event
window return}. \end{array} \right. \end{aligned}$$
GSARs are homogeneous in that, under the null hypothesis of no event
(mean) effect, they have zero mean and unit variance. Using these results,
rank transformations can be defined as
$$\begin{aligned} U_{it} = \text {rank}( GSAR _{it}) / (T + 1) -
(A11.3)
1/2 \end{aligned}$$
for $$i = 1, \ldots , n$$, where T indicates the number of estimation
window returns plus the event window SCAR observation so that T refers
to the rank number associated with $$SCAR _i^*(\tau _1, \tau _2)$$.
As such, these ranks capture both single day abnormal returns as well as
cumulative abnormal returns. For testing the null hypothesis of no event
effect, the authors derived the following rank test statistic:
$$\begin{aligned} t_{\textrm{grank}} = Z\left( \frac{T - 2}{T - 1
(A11.4)
- Z^2}\right) ^{\frac{1}{2}}, \end{aligned}$$
where
$$\begin{aligned} Z = \frac{\bar{U}_{T}}{s_{\bar{U}}}
\end{aligned}$$
with
$$\begin{aligned} s_{\bar{U}} = \sqrt{\frac{1}{T} \sum _{t = 1}^T
\frac{n_t}{n} \bar{U}_{t}^2}, \end{aligned}$$
$$\begin{aligned} \bar{U}_t = \frac{1}{n_t} \sum _{i = 1}^{n_t}U_{it},
\end{aligned}$$
and $$n_t$$ is the number of available cross-section returns at time
$$t = 1, \ldots , T$$.
Under the assumption of cross-sectional independence, the asymptotic
null distribution of $$t_{\textrm{grank}}$$ is Student’s t-distribution
with $$T - 2$$ degrees of freedom. In addition to event-induced
variance, $$t_{\textrm{grank}}$$ accounts for cross-sectional
correlation of abnormal returns when the event days are clustered.
Similar to Kolari et al. (2018), Pynnonen (2022) developed an adjustment
for $$t_{\textrm{grank}}$$ to account for cross-sectional
dependence in cases where the event windows are partially clustered.
The major attractiveness of nonparametric tests is their robustness to
non-normality. Also, rank tests are invariant to the usage of simple
returns or log-returns as log-transformation preserves the order of
observations. Many researchers conduct both nonparametric and parametric
tests to check the robustness of the results.

References
Aharony, J., and I. Swary. 1980. Quarterly dividend and earnings announcements and stockholders’
returns: An empirical analysis. Journal of Finance 35: 1–12.
[Crossref]

Ashley, J.W. 1962. Stock prices and changes in earnings and dividends. Journal of Political Economy
70: 82–85.
[Crossref]

Baker, C.A. 1956. Effective stock splits. Harvard Business Review 34: 101–106.

Baker, C.A. 1957. Stock splits in a bull market. Harvard Business Review 35: 72–79.

Baker, C.A. 1958. Evaluation of stock dividends. Harvard Business Review 36: 99–114.

Ball, R., and P. Brown. 1968. An empirical analysis of accounting income numbers. Journal of
Accounting Research 6: 159–178.
[Crossref]

Barber, B.M., and J.D. Lyon. 1997. Detecting long-run abnormal stock returns: The empirical power
and specification of test statistics. Journal of Financial Economics 43: 341–372.
[Crossref]

Binder, J.J. 1985. On the use of the multivariate regression model in the event studies. Journal of
Accounting Research 23: 370–383.
[Crossref]
Binder, J.J. 1998. The event study methodology since 1969. Review of Quantitative Finance and
Accounting 11: 111–137.
[Crossref]

Boehme, R.D., and S.M. Sorescu. 2002. The long-run performance following dividend initiations and
resumption: Underreaction or product chance? Journal of Finance 57: 871–900.
[Crossref]

Boehmer, E., J. Musumeci, and A.B. Poulsen. 1991. Event study methodology under conditions of
event induced variance. Journal of Financial Economics 30: 253–272.
[Crossref]

Brockett, P.L., H.-M. Chen, and J.R. Graven. 1999. A new stochastically flexible event methodology
with application to Proposition 103. Insurance: Mathematics and Economics 25: 197–217.

Brown, S.J., and J.B. Warner. 1980. Measuring security price performance. Journal of Financial
Economics 8: 205–258.

Brown, S.J., and J.B. Warner. 1985. Using daily stock returns: The case of event studies. Journal of
Financial Economics 14: 3–31.
[Crossref]

Cameron, A.C., J.B. Gelbach, and D.L. Miller. 2011. Robust inference with multiway clustering.
Journal of Business and Economic Statistics 29: 238–249.
[Crossref]

Campbell, J.Y., A.W. Lo, and A.C. MacKinlay. 1997. The Econometrics of Financial Markets.
Princeton, NJ: Princeton University Press.
[Crossref]

Corrado, C. J. 1989. A Nonparametric test for abnormal security-price performance in event studies.
Journal of Financial Economics 23: 385–395.
[Crossref]

Dolley, J.C. 1933. Characteristics and procedure of common stock split-ups. Harvard Business
Review 11: 316–326.

Dutta, A., J. Knif, J.W. Kolari, and S. Pynnonen. 2018. A robust and powerful test of abnormal stock
returns in long-horizon event studies. Journal of Empirical Finance 47: 1–24.
[Crossref]

Fama, E.F. 1998. Market efficiency, long-term returns, and behavioral finance. Journal of Financial
Economics 49: 283–306.
[Crossref]

Fama, E.F., and K.R. French. 1993. Common risk factors in the returns on stocks and bonds. Journal
of Financial Economics 33: 3–56.
[Crossref]

Fama, E.F., and K.R. French. 2015. A five-factor asset pricing model. Journal of Financial
Economics 116: 1–22.
[Crossref]
Fama, E.F., L. Fisher, M.C. Jensen, and R. Roll. 1969. The adjustment of stock prices to new
information. International Economic Review 10: 1–21.
[Crossref]

Harrington, S.E., and D.G. Shrider. 2007. All events induce variance: Analyzing abnormal returns
when effect vary across firms. Journal of Financial and Quantitative Analysis 42: 229–256.
[Crossref]

Ikenberry, D., J. Lakonishok, and T. Vermaelen. 1995. Market underreaction to open market share
repurchases. Journal of Financial Economics 39: 181–208.
[Crossref]

Jaffe, J.F. 1974. The effect of regulatory changes on insider trading. Bell Journal of Economics and
Management Science 5: 93–121.

Knif, J., J.W. Kolari, and S. Pynnonen. 2008. Stock market reaction to good and bad inflation news.
Journal of Financial Research 31: 141–166.
[Crossref]

Kolari, J.W., and S. Pynnönen. 2010. Event study testing with cross-sectional correlation of abnormal
returns. Review of Financial Studies 23: 3996–4025.
[Crossref]

Kolari, J.W., and S. Pynnönen. 2011. Nonparametric rank tests for event studies. Journal of
Empirical Finance 18: 953–971.
[Crossref]

Kolari, J.W., B. Pape, and S. Pynnonen. 2018. Event study testing with cross-sectional correlation
due to partially overlapping event windows. Mays Business School Research Paper No. 3167271.
Available at SSRN: https://ssrn.com/abstract=3167271 or https://doi.org/10.2139/ssrn.3167271.

Kolari, J.W., S. Pynnonen, and A.M. Tuncez. 2021. Further evidence on long-run abnormal returns
after corporate events. Quarterly Review of Economics and Finance 81: 421–439.
[Crossref]

Kothari, S.P., and J.B. Warner. 2006. Econometrics of event studies. In Handbook of corporate
finance: Empirical corporate finance, Vol. 1, ed. B.E. Eckbo, Chapter 1, 1–36. Amsterdam: North
Holland.

Lyon, J.D., B.M. Barber, and C.-L. Tsai. 1999. Improved methods for tests of long-horizon abnormal
stock returns. Journal of Finance 54: 165–201.
[Crossref]

MacKinlay, C.A. 1997. Event studies in economics and finance. Journal of Economic Literature 35:
13–39.

Myers, J.H., and A.J. Bakay. 1948. Influence of stock split-ups on market price. Harvard Business
Review 26: 251–265.

Patell, J. 1976. Corporate forecasts of earnings per share and stock price behavior: Empirical tests.
Journal of Accounting Research 14: 246–276.
[Crossref]

Pynnonen, S. 2022. Non-parametric statistic for testing cumulative abnormal stock returns. Journal
of Risk and Financial Management 15: 149. https://doi.org/10.3390/jrfm15040149.

Schwert, G.W. 1981a. Using financial data to measure the effect of regulation. Journal of Law and
Economics 24: 121–157.
[Crossref]

Schwert, G.W. 1981b. The adjustment of stock prices to information about inflation. Journal of
Finance 36: 15–29.
[Crossref]

Schipper, K., and R. Thompson. 1983. The impact of merger-related regulations on the shareholders
of the acquiring firms. Journal of Accounting Research 21: 184–121.
[Crossref]

Sharpe, S.A. 2001. Reexamining stock valuation and inflation: The impact of analysts’ earnings
forecast. Review of Economics and Statistics 84: 632–648.
[Crossref]

Sunder, S. 1973. Relationship between accounting changes and stock prices: Problem of
measurement and some empirical evidence. Journal of Accounting and Research 11: 1–45.
[Crossref]

Sunder, S. 1975. Stock prices and related accounting changes in inventory valuation. Accounting
Review 50: 305–315.

Footnotes
1 See, for example, Dolley (1933), Myers and Baker (1948), Baker (1956, 1957, 1958), Ashley
(1962), Ball and Brown (1968), and many others.

2 For example, see studies by Sunder (1973, 1975), Aharony and Swary (1980), Binder (1985),
and others.

3 For example, see studies by Schwert (1981a, b), Schipper and Thompson (1983), Brockett et al.
(1999), Sharpe (2001), Knif et al. (2008), and others.

4 See also Campbell et al. (1997) and Binder (1998).

5 Here the conditioning information set , and the conditional expectation itself is
simplified to the linear model $$E(R_{it}| R_{mt}) = \alpha _i + \beta _i R_{mt}$$, where the
coefficients are assumed to be time independent.

6 In statistical testing, Type I error means falsely rejecting the null hypothesis when it is true (false
alarm), and Type II error means accepting the null hypothesis when it is not true (missed alarm). The
probability of Type I errors is set by the researcher. Typical values are 5% or 1%. The Type II error
probability depends on the Type I error probability, sample size, distribution of the sample statistic,
and the extent to which the true parameter value deviates from the null hypothesis value. Statistical
power is measured by $$1 - \text {Type-II-probability}$$, or the probability of detecting a false
null hypothesis (correct alarm). It is important to use statistical tests that have maximum power in
event studies.

7 The respective standard errors in Eqs. (11.7) and (11.9) are defined as:
$$\begin{aligned} \mathrm {s.e}({\overline{ AR }_0}) = \sqrt{\frac{1}{n(n - 1)} \sum _{i =
1}^n\left( AR _{i0} - \overline{ AR }_0\right) ^2} \qquad {(11.11)} \end{aligned}$$
and
$$\begin{aligned} \mathrm {s.e}({\overline{ CAR }_{\tau _1, \tau _2}}) = \sqrt{\frac{1}{n(n-1)}
\sum _{i = 1}^n\left( CAR _i(\tau _1, \tau _2) - \overline{ CAR }_{\tau _1, \tau _2}\right) ^2}.
\qquad {(11.12)} \end{aligned}$$

8 The standard error used in the CAR t-statistic (11.9) is an example of clustering robust standard
errors (with respect to serial correlation), where the event windows over which individual CARs are
aggregated from the clusters. See Cameron et al. (2011), Dutta et al. (2018), and Kolari et al. (2018)
for further information.

9 With the market model $$R_{it} = \alpha _i + \beta R_{mt} + e_{it}$$, $$d_{it}$$ in
Eq. (11.14) becomes:
$$\begin{aligned} d_{it} = \frac{1}{T} + \frac{(R_{mt} - \bar{R}_m)^2}{\sum _{s = 1}^{T}
(R_{ms} - \bar{R}_m)^2} \end{aligned}$$
and $$d_{i\tau }$$ in Eq. (11.16)
$$\begin{aligned} d_{i\tau } = \tau ^2\left( \frac{1}{T} + \frac{(\bar{R}_{m\tau } -
\bar{R}_m)^2}{\sum _{s = 1}^{T}(R_{ms} - \bar{R}_m)^2}\right) , \end{aligned}$$
where T is the estimation window length, $$\bar{R}_m$$ is the average market return in the
estimation window, and $$\bar{R}_{m\tau }$$ is the average market return in the window over
which CAR is computed. Because calendar times of the event are assumed not overlapping,
estimation and event windows are unique for each stock i, such that we use the subscript i in
$$d_{it}$$ and $$d_{i\tau }$$.
 In general, for a factor model with p factors, the correction terms are
$$d_{it} = x_t' (X'X)^{-1}x_t$$ and
$$d_{i\tau } = \tau ^2\bar{x}_\tau ' (X'X)^{-1}\bar{x}_{\tau }$$, where
$$x_t = (1, F_{1t}, \ldots , F_{pt})'$$ includes event time t returns,
$$\bar{x}_{\tau } = (1, \bar{F}_{1\tau }, \ldots , \bar{F}_{p\tau })'$$ includes factor averages
over the CAR-window of length $$\tau$$, and $$(X'X)^{-1}$$ is the inverse of the
$$(p+1)\times (p+1)$$ matrix of estimation window cross-products of the constant term and
factor returns.

10 See Campbell et al. (1997, Chapter 4) for an excellent discussion and further details.

11 Approaches for taking into account cross-sectional correlation with clustering robust estimation
methods as well as estimating the average cross-sectional correlation explicitly in the partial
clustering case are discussed in Kolari et al. (2018).

12 USA Today (September 8, 2013): https://eu.usatoday.com/story/money/business/2013/09/08/

chronology-2008-financial-crisis-lehman/2779515/COVID-19.
AJMC: https://www.ajmc.com/view/a-timeline-of-covid19-developments-in-2020.

13 https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html.

14 This standard error is computed as:

$$\begin{aligned} \mathrm {s.e}(\overline{ BHAR }) = \sqrt{\frac{1}{n(n-1)}\sum _{i = 1}^n(
BHAR _{iT} - \overline{ BHAR })^2}. \qquad {(11.25)} \end{aligned}$$
Assuming the independence of $$BHAR _{iT}$$s, the asymptotic null distribution of
$$t_{ BHAR }$$ is standard normal.

15 We should note that, because the number of stocks in each month can vary from 1 to n (the total
number of stocks), weighted least squares are recommended in the regression estimation.
Index
A
Abnormal return (AR) 19, 22, 157, 201–209, 211, 214–219
Accounting variables 165–167
Achilles’ heel 134, 215
Actively managed mutual funds 45, 157
Actual return 187, 188, 198, 203, 205
Actuarial value 23, 34, 36
Adjusted BMP test 208
Adjusted R-squared 67, 74
Adler, Michael 113, 114, 123, 125
Aggregate consumption 116, 117, 119, 121, 124
Aggregate consumption beta 121
Aggregate weights 121
Aharony, J. 202
Alpha term 121, 125
Anomaly 21, 92, 147, 156, 158, 162, 163
Arbitrage 19, 20, 40, 92, 129, 130, 136, 157, 168, 195
Arbitrage portfolio 130–134
Arithmetic mean return 18, 35
Artificial intelligence (AI) 155, 169, 172
Ashley, J.W. 201
Asset pricing anomalies 150, 160, 161, 172
Asset pricing theory (APT model) 130, 132–134, 146, 150
Average market return 61, 182, 184
Average portfolio return 5, 8, 10
Average return 5, 7–9, 11, 62, 91, 97, 140, 141, 161, 166, 182, 184, 186,
217
Axis of symmetry of the parabola 180, 196
B
Back, K. 165
Bad model problem 92
Bakay, A.J. 201
Baker, C.A. 201
Ball, R. 201
Banz, R.W. 140
Barber, B.M. 215, 218
Barberis, M. 92
Barillas, F. 161, 163, 164
Barroso, P. 158
Behavioralists 19, 21, 22, 34, 92, 97, 216
Benchmark asset pricing model 214
Best linear unbiased estimator (BLUE) 69–71, 73, 75
Beta 47
Beta coefficients 71, 91, 92, 108, 124, 141, 143, 152, 165
Beta loading 119, 136, 141, 167
Beta risk 47, 50–54, 57–59, 61–63, 67, 75, 86–88, 91, 92, 98, 108, 111,
120, 129, 139, 148–150, 166, 171, 177, 179, 180, 182–184, 186, 187, 194–
196
Beta risk loadings 141, 148
Bhandari, L.C. 140
Binder, J.J. 202
Black, F. 51, 58, 59, 61, 64, 75, 85–87, 90, 96, 107, 111, 141, 147, 148,
151, 177, 179, 195, 196
Blume, M. 60, 62, 90
BMP test 207
Boehme, R.D. 216
Boehmer, E. 207, 208, 218
Bond market model 144
Book equity to market equity ratio (BM) 91, 140
Breeden, N.D. 115–117, 123–125
Brockett, P.L. 202
Brownian motion 106
Brown, P. 201
Brown, S. 202, 204
Butt, H.A. 177, 195
Buy-and-hold abnormal return (BHAR) 214, 215, 218
C
Calendar time abnormal return (CTAR) 214–216, 219
Cameron, A.C. 71, 76, 80, 206
Campbell, J.Y. 202, 208
Capital Asset Pricing Model (CAPM) 26, 39, 42, 57
Capital investment factor 159, 163, 171
Capital Market Line (CML) 42, 44, 53, 54, 100
Carhart, M.M. 155–158, 161, 163, 171, 177, 187, 192, 196, 198
Center for Research in Security Prices (CRSP) 58, 93, 117, 142, 146, 172,
180, 186, 187, 217
Central Limit Theorem (CLT) 29, 35, 36, 80, 206, 221
Certain return 23, 34
Certainty equivalent 23, 50
Certainty equivalent valuation formula 50, 53
Chen, H.-M. 202
Chen, N.-F. 136, 137
Chordia, T. 193
Clustered event days 209, 223
Clustering robust standard errors 71, 80
CMA factor 159
Cochrane, J.H. 41, 63
Coefficient of determination 73
Common factor 129–131, 133, 134, 146, 147
Computer algorithms 15, 168, 169, 172
Conditional CAPM 105, 109, 120–122, 124
Conditional mathematical model 166, 167, 172
Conditional model 119, 165, 166, 172
Conditional state variables 124
Consumption CAPM (CCAPM) 105, 115, 122, 123
Consumption pattern 115
Continuously compounded return 27, 28, 35
Continuous-time process of asset pricing 106
Control firm 215–217
Copeland, Thomas 49
Corrado, Charles 222
Correlation of returns 4, 46
Covariance of returns 4
2020 COVID-19 pandemic 210, 218
Cross-sectional correlation 208, 209, 223
Cross-sectional regression 61–63, 75, 89, 90, 96, 119, 135, 136, 141, 148,
164, 165, 168, 172, 186, 199
Cross-sectional return 9
Cross-sectional standard deviation of returns 177, 195, 197
Cross-section factor model 166, 169
Cross-section factors 164–168, 172
Cumulative abnormal return (CAR) 203, 209, 211, 218, 222, 223
Currency risk 113, 123
D
Daniel, K. 92, 149, 151, 158
Data dredging exercise 164
Data mining 147, 151
De bondt, W.F.M. 92
Decile portfolios 140
Default risk (DEF) 145
Demand side theory 118
Dempster, A.P. 184
Dependent variable 67–69, 71, 73, 76, 80, 143, 148, 151, 158
Dhrymes, P. 134
Discount rate 2, 11, 40, 41, 50, 52, 159
Discretionary methods 169, 170, 172, 178
Diversifiable unsystematic risk 53
Diversification 1, 6, 7, 9–11, 33, 35, 53, 134, 139, 169
Diversified portfolio 45, 133, 136
Dividend discount model 159, 164
Dolley, J.C. 201
Domestic factors 111
Domestic model 112
Douglas, G.W. 62, 90
Dumas, Bernard 113, 114, 123, 125
Dummy signal variable 184, 196
Dutta, A. 206
Dybvig, P. 134
E
Economic significance 203, 218
Economy-wide events 202
Efficient frontier 8, 9, 11, 42–44, 46, 50, 53, 55, 87, 88, 93–96, 179–181
Efficient market hypothesis (EMH) 20, 34, 92, 97
Efficient portfolio 8, 10, 11, 44, 50, 52, 53, 87, 88, 91, 96, 97, 99, 100, 179
Efficient portfolio I 14, 88
Empirical model 58, 97, 108, 147, 195, 199
Empirical ZCAPM 184–191, 194, 196–198
Endogeneity problem 72, 76, 149, 151
Endogenous portfolio 190
Equilibrium prices 19, 20, 26
Equilibrium relationship between risk and return 51
Error-in-variables 62
Estimation window 63, 185, 203, 204, 207, 208, 211, 212, 218, 222, 223
Euler, Leonard 28
Event 202
Event day 203, 205, 207–209, 211, 212
Event-induce variance 206, 208, 222, 223
Event study 201, 202, 209, 211, 214
Event window 203–207, 209, 211–213, 215, 218, 222, 223
Ex ante return 11, 42, 49
Exchange rate movements 112, 113, 123
Exchange rate puzzle 123
Exchange rate risk 110, 112, 113, 123
Exchange risk exposure coefficient 115
Exogenous portfolio 151
Expectation-maximization (EM) regression 184, 196, 199
Expected lifetime utility 115, 123
Expected rate of return 2, 11, 41, 42, 46, 47, 53, 58, 64
Expected return 2–4, 10, 11, 13, 17, 20, 22, 24, 44, 45, 47, 63, 86, 87, 91,
92, 99, 100, 106, 108, 116–118, 129, 130, 132, 133, 136, 140, 146–150,
179, 180, 184, 203
Expected return of a portfolio 3, 44
Expected utility of wealth 42, 86
Explanatory variable 67–69, 71, 72, 74, 135
Ex post return 4, 11, 41, 49
F
Factor loading 89, 90, 135, 136, 158, 164, 168, 172, 177, 192
Fama, E.F. 21, 34, 50, 58, 61, 62, 75, 88, 90, 92, 96, 117, 135, 141, 148,
164, 167, 172, 177, 188, 192, 194, 196, 201, 202, 204
Ferson, W.E. 89, 91, 120, 167
2008 financial crisis 210, 213, 218
Firm-specific events 202
Fisher, L. 201, 202, 204
Five-factor model 119, 159–163, 165, 168, 171, 172, 204, 211
Four-factor model 155, 156, 158–162, 171, 172, 177, 187, 192
Friedman, M. 22, 23, 34
Friend, I. 60, 62, 90, 134
F-statistic 91, 94, 96
Fundamental economic forces 52
G
Gauss-Markov Theorem 69, 75
Gelbach, J.B. 71, 76, 80, 206
General equilibrium conditions 195
Generalized standardized abnormal return (GSAR) 222
Geometric mean return 18, 36
Gibbons, M.R. 62, 117, 120, 124, 125
Gibbons, Ross, and Shanken (GRS) test 159, 186
Global factors 111
Global minimum variance portfolio G 8, 9, 87
Global model 112
Goodness-of-fit 57, 67, 69, 73
Gordon, M.J. 159
Goyal, A. 193
Graven, J.R. 202
Green, R.C. 88
Griffin, J.M. 111, 123
Gross log return 27
Gross return 18, 27, 28, 157
Gultekin, N.B. 134
H
Hansen, L.P. 120
Harrington, S.E. 206
Harvey, C.R. 120, 192
Hedge portfolio 108, 111, 123, 195
Heteroskedasticity 71, 82
Hidden variable 184
Hirshleifer, D. 92
HML factor 142, 144, 146, 162
Hodrick, R.J. 120
Hollifield, B. 88
Homogeneous beliefs 20, 130
Homoskedasticity 69, 71, 75
Hou, K. 155, 160–162, 171, 172, 178, 187, 194, 196
Huang, J. 9, 99, 148, 177, 182, 184, 192, 194, 195
Huang, J.Z. 177, 195
Huberman, G. 129, 132, 133
I
Idiosyncratic risk 47, 52, 130, 131, 134, 136
Ikenberry, D. 215, 218
Independent variable 62, 68, 69, 71, 74–76, 80, 94, 143, 147–149, 151
Index funds 45, 171
Indifference curve 17, 24, 26, 34, 42, 43, 45
Inefficient portfolio ZI 9
Industry portfolios 95, 148, 149, 151, 159, 178, 190, 192, 195, 196, 210
Inflation rate 113
Ingersoll, J.E. 133
Instrumental variable 72, 76, 124
Integrated world market 109, 123
Interaction terms 166
Intercept term 57, 58, 62, 64, 68, 69, 71, 75, 76, 79, 89, 96, 111, 117, 123,
165
International asset pricing model (IAPM) 109, 123
International Monetary Fund (IMF) 114
Intertemporal CAPM 91, 105–107, 123, 168
Investment-based model 119
Investment/capital ratio 119, 124
Investment expenses 157
Investment opportunity set 46, 106, 107, 109, 111, 120, 123
Investments to assets ratio 160, 194
I* portfolio 181, 199
Irrational investor behavior 92, 97
J
Jaffe, J.F. 216, 219
Jagannathan, R. 88, 141
Jegadeesh, N. 21, 156
Jensen’s alpha 59, 69, 74, 90, 96
Jensen, M.C. 58, 59, 75, 86, 90, 111, 141, 201, 202, 204
Jones, P.N. 184
Jorion 114, 123, 125
K
Kallberg, J.G. 88
Kandel, S. 120
Kapadia, N. 165
Knif, J. 202, 206
Kolari, J.W. 9, 99, 106, 148, 177, 182, 184, 192, 194, 195, 202, 206, 208,
209, 218, 222, 223
Kothari, S.P. 148, 149, 151, 152
Kozak, S. 169
Krishnan, T. 185
Kroll, Y. 88
Kurtosis 18, 31, 32, 34, 35
L
Lagged variable 121
Laird, N.M. 184
Lakonishok, J. 92, 140, 215, 218
Lanstein, R. 140
Latent factor 169, 172
Latent factor model 169
Lettau, M. 121, 122, 124, 155, 167–169, 172
Levy, H. 88
Levy, M. 88
Lewellen, J. 121, 122, 124, 149, 151
Liao, H. 177, 195
Lifetime consumption 106, 109
Linear relation 67, 92, 129, 134, 136, 151
Linnainmaa, J. 193
Lintner, J. 26, 39, 42, 50–53, 57, 139, 177, 195
Litzenberger, R.H. 117, 124, 125
Liu, W. 9, 99, 148, 177, 182, 184, 192, 194, 195
Liu, Y. 192
Load fees 157
Loadings 62, 63, 75, 135, 141, 165
Lo, A.W. 202, 208
Log return 18, 27–30, 34, 35
Long position 86, 88
Long-run event studies 213, 214, 217, 218
Long/short portfolio 89, 136, 142, 164, 178, 196
Lucas, R.E. 115–117, 123
Ludvigson, S. 121, 122, 124
Lyon, J.D. 215, 218
M
MacBeth, J.D. 61, 62, 75, 88, 90, 96, 117, 135, 141, 148, 164, 167, 172,
177, 188, 192, 194, 196
Machine learning methods 155, 169, 171, 172, 178, 197
Machine learning technique 168
MacKinlay, A.C. 202, 208
MacKinlay, C.A. 202
Malkiel, B.G. 21, 34
Management expenses 157
Management factor 161
Man versus machine 172
Marginal rate of substitution 40, 115
Market anomaly 98
Market capitalization 58, 64, 91, 140, 141, 144, 145, 160, 186, 217
Market factor 52, 57, 61, 67, 73, 90–92, 96, 105–108, 111, 118, 120, 121,
123, 136, 139, 141, 142, 144–146, 149, 150, 163, 167, 171
Market imperfections 20, 52
Market model 57–60, 62, 64, 66, 67, 71–76, 86, 88, 90, 96, 109–111, 121,
140, 201
Market portfolio 44–47, 50, 52, 53, 55, 58, 86, 109, 117, 129, 133, 134,
149
Market portfolio proxy 60, 62–64, 74, 75
Market price of risk 62, 63, 75, 123, 141, 164
Markowitz, H.M. 1, 3–8, 10, 11, 13, 17, 18, 20, 22, 24, 26, 39, 42, 50, 53,
58, 62, 74, 88. 90, 96, 150, 177, 195
Ma, T. 88
Matched control firm 216
Mathematical model 165, 172
Matlab program 186
McGrattan, E.R. 141
McLachlan, G.J. 184, 185
McLean, R.D. 193
Mean model 203, 204, 208, 211
Mean-variance parabola 1, 8, 96, 178, 180, 183, 184, 196
Merton, R.C. 91, 97, 100, 105–110, 115, 120, 123, 146, 150, 168, 195
MGMT factor 161
Michaud, R.O. 10
Miller, D.L. 71, 76, 80, 206
Miller, M. 62, 90, 135
Mimicking portfolio 142, 164, 165, 172
Minimum variance portfolios 8, 9, 11, 179
Minimum variance unbiased estimator (MVUE) 71
Mispricing alpha 69, 165
Mispricing error 67, 145, 157, 158, 162, 168, 169, 216
Missing factors 59, 71, 73–75, 89, 90, 96, 111
Model parameters 75, 78, 167
Momentum factor 122, 155–158, 160, 164, 169, 171, 172
Momentum profits 21, 156, 157
MOM factor 163
Morgenstern, O. 22, 34
Moskowitz, T. 158
Mossin, J. 26, 42, 50, 51, 53, 57, 139, 177, 195
M-talk 41
Multicollinearity 71, 144
Multifactor APT model 146, 150
Multifactors 108, 123, 142, 148, 151, 155, 159, 160, 164, 165, 169, 171,
172, 178, 195, 197
Multilateral Exchange Rate Model (MERM) 114
Multinational firms 114, 115, 123
Multiple regression 68–71, 75, 108
Musumeci, J. 207, 208, 218
Mutual funds 156, 157, 171
Myers, J.H. 201
N
Nagel, S. 121, 122, 124, 149, 151, 169
Nallareddy, S.K. 167
Napier, John 28
Negative return dispersion 184, 196
Newey, W. 71, 75, 80
No-arbitrage condition 130, 132, 136
Nobel, Alfred 139
No load mutual fund 157
Nonlinear relation 62
Nonparametric test 204, 221, 223
Nonresidential investments 119
Normal distribution 29–32, 78–80, 85, 204, 221
Normality of returns 130, 221
Normative models 62, 188
Novy-Marx, R. 159
Null hypothesis 71, 79, 80, 93, 94, 205, 206, 215, 216
O
Optimal portfolio weights 15
Ordinary least squares (OLS) regression 57, 58, 184
Orthogonal portfolios 180, 181, 196
Ostdiek, B. 165
Out-of-sample return 196
Overlapping calendar days 209
Overpriced asset 130
P
Pape, B. 206, 209, 223
Parameter 30, 53, 57, 63, 67–69, 73, 76, 105, 124, 192
Parametric test 204, 205, 223
Parsimonious factor model 163
Partial clustering 209
Passively managed mutual funds 157
Patell, J. 207, 208, 218
Patell test 209
Peel, D. 184
Pelger, M. 155, 167–169, 172
Perfect capital market 17–20, 39, 52
Perfect certainty 20, 34
Performance factor 161
Pontiff, J. 193
Population mean 10, 205
Portfolio return 4–6, 10–12, 18, 62, 64, 75, 89, 90, 110, 143, 158, 162, 178,
216
Positive return dispersion 178, 185, 196
Poulsen, A.B. 207, 208, 218
Power 74, 113, 151, 161, 168, 207, 217
Predicted return 188, 203, 205, 214, 218
Present value formula 1, 11, 40, 41, 52
Present value of future cash flows 40
Price taker 20, 34
Principal Component Analysis (PCA) 167
Probability 22, 23, 41, 42, 131, 196
Probability of positive (or negative) return dispersion effect 185
Production CAPM (PCAPM) 105, 118, 124, 171
Profit factor 159, 171
Pulley, L.B. 88
Pynnonen, S. 202, 206, 208, 209, 218, 222, 223
Python program 58, 186
Q
Quadratic utility function 26, 42, 129, 130
q-theory 160
R
Random Walk Hypothesis (RWH) 17, 21, 34, 201
Rank test 223
Realized return 4, 20
Regression analysis 57, 58, 64, 73
Regression coefficient 57, 68–71, 75, 114, 165, 166, 196, 207
Reid, K. 140
Residential gross investments 119
Residual error 62, 149, 201
Return dispersion 148, 177, 178, 182, 183, 186, 194, 195
Return on equity 160, 194
Rise of the machines 169
Risk-adjusted rate of return valuation formula 49, 50, 53
Risk-adjustment 92, 214, 215
Risk averse 2, 9, 23, 24, 34
Risk dimension 41, 108, 135, 167
Riskless asset 42, 44, 46, 50, 86, 87, 107, 110, 133
Riskless rate 2, 11, 41, 42, 44, 47, 50, 52, 53, 57, 72, 85–87, 90, 96, 97,
105–108, 110, 116, 117, 123, 130, 133, 187, 196, 216
Risk loving 22, 24, 34
Risk neutral 22
Risk preference 9, 23, 50, 53, 57, 110
Risk premium 2, 11, 17, 23, 34, 41, 42, 50, 52, 63, 64, 89, 96, 110, 111,
115, 116, 124, 135, 143, 159
Ritov, Y. 88
RMW factor 159, 165
Roberts, M. 193
Robustness 82, 164, 204, 208, 223
Roll critique 59
Roll, R. 134–136, 201, 202, 204
Rosenberg, B. 140
Ross, S.A. 92, 97, 134–136
Rough measures of return dispersion 195, 197
R program 58, 186, 211
R squared 67, 73, 74, 76
Rubin, D.B. 184
S
Sample mean return 135
Sample variance of returns 5
Santa-Clara, P. 158
Santosh, S. 169
Saretto, A. 193
SAS regression software 211
Savage, L. 22, 23, 34
Scaling abnormal returns 201
Schipper, K. 202
Scholes, M. 58, 59, 62, 75, 86, 90, 111, 135, 141
Schwert, G.W. 202
Seasoned equity offering (SEO) 209, 217–219
Security Market Line (SML) 47, 53, 57, 59, 75, 85, 96
Semi-strong form efficiency 21
Shanken, J. 62, 63, 148, 149, 151, 152, 161, 163, 164
Shapiro, A.C. 140
Sharpe, W.F. 26, 39, 42, 46, 57, 74, 97, 139, 177, 195
Shleifer, A. 92
Short-run event studies 202, 204, 209, 214, 218
Short sell 42
Shrider, D.G. 206
Signal variable 184, 185
Simin, T. 122, 167
Simple return 18, 26, 29, 34, 35, 223
Six-factor model 155, 161, 164, 166, 172, 178, 187, 191, 192, 196
Size factor 142, 143, 147, 162–164, 172, 178
Skewness 18, 31, 32, 34, 35, 222
Sloan, R.G. 148, 149, 151, 152
SMB factor 142, 143, 146
Smooth consumption 116
Solnik, B.H. 109–111, 123, 125
Sorescu, S.M. 216
Sorting 60, 163, 187
Span of the parabola 179, 197
Special Drawing Right (SDR) 114
Stable distributions 32
Stambaugh, R.F. 120, 155, 161–163, 171, 178, 187, 194, 196
Standard deviation of returns 3, 5, 7, 11, 17, 42, 44–46, 53, 54, 135
Standard deviation of returns in a portfolio 46
Standard error 60, 63, 71, 76, 79, 80, 205, 206, 215, 222
Standardized abnormal returns 218, 222
Standardized cumulative abnormal return (SCAR) 222
Stata regression software 211
State variable 91, 92, 106–108, 111, 120, 122, 123, 147, 151, 159
Statistical significance 192, 207
Statman, D. 140
Stochastic discount factor 40
Stock splits 27, 64, 201, 202
Strong-form efficiency 21, 34
Subrahmanyam, A. 92
Sunder, S. 202
Supply side theory 118
Survivor bias 148, 151
Swary, I. 202
Systematic risk 47, 52–54, 57, 67, 89, 108, 116, 122, 124, 131, 136, 183,
184, 201
T
Tangent portfolio 50
Taxes 19, 52, 106
Taylor series 30, 121
Technical analysis 21
Term-to-maturity (TERM) 144
Test asset portfolios 88, 144, 148, 149, 160, 161, 178, 186, 187, 192, 194
Test assets 90, 119, 143, 148, 149, 151, 156, 159, 160, 162, 195
Thaler, R.H. 92
Theoretical methods 171, 172
Theoretical (ZCAPM) 171
Thompson, R. 202
Three-factor model 91, 92, 97, 121, 122, 139, 140, 142–147, 149–151, 155,
156, 158, 160, 161, 163, 171, 190, 216
Three-fund theorem 107, 109, 123
Time-series regression 63, 64, 75, 89, 91–93, 96, 111, 118, 123, 135, 147,
150, 158, 186
Time-varying loadings 165
Titman, S. 21, 149, 151, 156
Tobin, J. 24, 26, 34, 39, 42, 44, 50
Total risk 3, 6, 7, 42, 44, 46, 47, 53
Transactions costs 10, 19, 20, 34, 45, 52
Treynor, J.L. 26, 39, 42, 50, 51, 53, 57, 139, 177, 195
Tsai, C.-L. 215, 218
t-statistic 61, 63, 67, 78, 80, 111, 141, 143, 148, 193–196, 205–207
t-test 60, 61, 63, 93, 218
Turnover 157
Two-factor model 96, 108, 111, 146, 196
Two-sided, opposite effects of return dispersion 182
U
Unconditional CAPM 120–122, 124
Uncorrelated 6, 30, 61, 64, 69, 96, 144, 149, 168
Underpriced asset 54, 130
Undiversifiable systematic risk 53
Unsystematic risk 47
Utility 17, 22–24, 26, 34, 42, 50, 115, 117
V
Value factor 91, 92, 97, 111, 121, 122, 142–144, 146, 148, 149, 151, 155,
162, 171, 190, 216
Value-weighted market return dispersion 186
Variance of portfolio returns 4–6, 10–12
Variance of returns 3, 6, 8, 11, 12, 17, 24, 26, 30, 34, 87, 178–181, 196
Vermaelen, T. 215, 218
Vishny, R. 92
Von Neumann, J. 22, 34
W
Warner, J.B. 202, 204
Weak-form efficiency 21
Wealth 18, 43, 51, 91, 117, 123, 147
West, K. 71, 75, 80
Weston, J. Fred 49
White, H. 71, 75, 80
Width of the parabola 9, 180
X
Xie, B. 167
Xue, C. 155, 160–162, 171, 172, 178, 187, 194, 196
Y
Yuan, Y. 155, 161–163, 171, 178, 187, 194, 196
Z
ZCAPM 97, 105, 148, 171, 172, 177, 178, 181, 182, 184, 185, 187, 190–
192, 194–196
Zero-beta CAPM 85, 86, 88–92, 96, 97, 99, 105, 106, 177, 179–182, 195,
196
Zero-beta portfolio 64, 86, 87, 89, 90, 96, 97, 100, 107, 117, 148, 179, 182,
196
Zero-investment portfolio 86, 89, 130, 133, 142, 146, 150, 162, 172, 212
Zeta coefficient 184, 188
Zeta risk 177, 178, 182–184, 186, 187, 195, 196
Zeta risk loadings 187, 194, 196, 198
Zhang, L. 155, 160–162, 171, 172, 178, 187, 194, 196
Zhu, H. 192
Ziemba, W.T. 88
ZI* portfolio 181, 199