ASSET PRICING THEORY

Claudio Tebaldi
ii
Contents

1 General Equilibrium Valuation. 1

1.1 The Capital Asset Pricing Model . . . . . . . . . . . . . . . . 1
1.1.1 The Market Portfolio . . . . . . . . . . . . . . . . . . . 2
1.1.2 The Capital Market Line . . . . . . . . . . . . . . . . . 2
1.1.3 The Security Market Line . . . . . . . . . . . . . . . . 2
1.1.4 SML and CML . . . . . . . . . . . . . . . . . . . . . . 3
1.1.5 CAPM and the present value of cash flows . . . . . . . 5
1.1.6 Risk Premium and diversification . . . . . . . . . . . . 6
1.2 The Zero-Beta CAPM (Black) . . . . . . . . . . . . . . . . . . 7
1.3 BAB (Frazzini and Pedersen, 2013) . . . . . . . . . . . . . . . 9

2 Partial Equilibrium Valuation 15

2.1 The linear factor model . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Arbitrage pricing theory . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Statement and Proof of the APT . . . . . . . . . . . . 19
2.2.2 Economic Interpretation of the λ coefficients . . . . . . 22
2.3 Traded portfolios as Risk factors . . . . . . . . . . . . . . . . . 22
2.4 Long-Short portfolios . . . . . . . . . . . . . . . . . . . . . . . 24
2.5 APT & CAPM . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 The pricing kernel 27

3.1 First order condition . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 The risk neutral measure . . . . . . . . . . . . . . . . . . . . . 27
3.3 APT and risk neutral valuation . . . . . . . . . . . . . . . . . 29
3.4 CAPM and SDF . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.5 Physical and risk-neutral probabilities . . . . . . . . . . . . . . 32
3.5.1 A numerical example . . . . . . . . . . . . . . . . . . . 34

4 Empirical tests 39
4.1 Empirical Test of Equilibrium Models . . . . . . . . . . . . . . 39
4.1.1 Ex-ante Expectations and Ex-post Tests . . . . . . . . 39

iii
iv CONTENTS

4.1.2 Some Hypotheses of the CAPM . . . . . . . . . . . . . 41

4.1.3 Common methodologies to test the CAPM . . . . . . . 41
4.1.4 An example of time-series test: Black, Jensen and Sc-
holes (1972) . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.5 Fama and MacBeth (1973) cross-sectional approach . . 45
4.1.6 A first example of cross-sectional test . . . . . . . . . . 46
4.1.7 The Roll’s Critique . . . . . . . . . . . . . . . . . . . . 49
4.2 Fama and French (1992) . . . . . . . . . . . . . . . . . . . . . 49
Chapter 1

General Equilibrium Valuation.

1.1 The Capital Asset Pricing Model

The formulation of the Capital Asset Pricing Model relies on the following
assumptions:

1. Mean Variance Portfolio Selection.

(a) Single Period Portfolio Selection.

(b) Agent preferences are consistent with the Mean Variance criterion.

2. Asset Markets are in equilibrium (prices can adjust so that the existing
stock of assets are willingly held)

(a) Frictionless and perfectly liquid markets:

i. zero transaction costs.
ii. no institutional restrictions to trades (short selling allowed).
(b) Ability to borrow unlimited amounts of money at the Rf .
(c) Asset are divisible into any desired unit.
(d) All assets can be bought / sold at the observed market price.
(e) Investors are price takers.
(f) Neutral Taxes (the same for every investor and every source of
income).

3. Homogeneous Beliefs and absence of information asymmetries.

1
2 CHAPTER 1. GENERAL EQUILIBRIUM VALUATION.

1.1.1 The Market Portfolio

The above assumptions imply that every investor will face the same efficient
frontier, in fact the risk free rate and the tangency portfolio will be the same
for all investors.
In equilibrium the supply equal demand condition implies that this tan-
gency portfolio has to be the Market Portfolio: the portfolio in which
every risky asset is represented with a weight equal to its share in the world
market capitalization. On the contrary, assume that some risky asset is not
present in the tangency portfolio, then no agent would demand that asset
and the asset would disappear from the market. Hence, under the CAPM
assumptions the investors do not choose to hold the market portfolio...they
choose to hold a tangency portfolio and since this portfolio is the same for
everybody, ”supply and demand” effects will make it the market portfolio.

1.1.2 The Capital Market Line

All the efficient portfolios are allocated along a linear frontier determined by
the following equation

µe − r0 µM − r0
=
σe σM
µM − r0
µe = r0 + σe
σM
The CML is the first representation of a market equilibrium that we
can derive under the CAPM assumptions. We see that the expected return
of a stock can be divided in three portions: the risk free rate plus a risk
premium. The ratio between the excess return and the risk feared by the
efficient portfolio is constant and equal to the same ratio for the market
portfolio. This is a good representation of the expected return and risk
trade-off for efficient portfolios. But what happens to inefficient portfolios?

1.1.3 The Security Market Line

In equilibrium it is possible to value the expected return also for an inefficient
portfolio thanks to the following:
Proposition 1.1 the Security Market Line describes a linear relation be-
tween the expected excess return of a generic portfolio and the excess return
of the market portfolio:
µi − r0 = βi (µM − r0 )
1.1. THE CAPITAL ASSET PRICING MODEL 3

portfolio are ranked according to the βi index, which is determined by the

covariance between the returns of the portfolio ri and the market portfolio rM

Cov (ri , rM )
βi = (1.1)
V ar (rM )

Proof. Let’s go back to the derivation of the MV frontier with the risky
assets and the risk free rate.
1 2 2
w1 σ1 + w22 σ22 + 2w1 w2 σ12 +λ (µp − r0 − w1 (µ1 − r0 ) − w2 (µ2 − r0 ))


min L =
{w,λ,γ} 2

Minimization implies (this result can be easily derived in the matrix solution):

(µ1 − r0 ) (µ2 − r0 ) µp − r0
= =
σ1,p σ2,p σp2
σi,p = Cov (ri , rP )

We can multiply everything for σp2

(µ1 − r0 ) (µ2 − r0 )
σ1p = σ2p = µp − r0
σp2 σp2

(µ1 − r0 ) (µ2 − r0 )
= = µp − r0
β1 β2
Since the market portfolio M is efficient, it is a minimum variance portfolio
hence by assuming p = M we get σσ1p2 = σσ1M2 = βi hence
p M

(µi − r0 )
= µM − r0
βi
µi = r0 + βi (µM − r0 )

The expected return of any inefficient portfolio is proportional to the β

and not to σ.
We can plot µi against βi and this is the SML.

1.1.4 SML and CML

It is important to understand that the SML is not an equilibrium condition
alternative to the CML: it is more general form of the same equilibrium.
We should remember that
4 CHAPTER 1. GENERAL EQUILIBRIUM VALUATION.

Cov (ri , rM )
βi = 2
σM
σi σM ρiM
= 2
σM
σi
= ρiM
σM

σi
SM L ⇒ rM ] − r0 )
µi = r0 + (E [e ρiM
σM
σe
CM L ⇒ rM ] − r0 )
µe = r0 + (E [e
σM

Now if we calculate the correlation coefficient between an efficient port-

folio and the market portfolio we get that

Cov (re , rM )
ρeM =
σM σe

Now since the efficient portfolio is a combination of market portfolio and

risk free we have that re = ωrM + (1 − ω) r0 and σe = ωσM .

Cov (re , rM ) = E [(re − µe ) (rM − µM )]

= E [(ωrM + (1 − ω) r0 − ωµM − (1 − ω) r0 ) (rM − µM )]
= E [(ωrM − ωµM ) (rM − µM )]
= E [ω (rM − µM ) (rM − µM )]
ωE (rM − µM )2
 
=
2
= ωσM

Going back to the correlation coefficient we see that

2
Cov (re , rM ) ωσM
ρe,M = = =1
σM σe ωσM σM

So, since for efficient portfolios the correlation coefficient with the market
portfolio is equal to one, we have that the CML is a special case of the SML
in case of perfect correlation where all the risk taken is compensated.
1.1. THE CAPITAL ASSET PRICING MODEL 5

1.1.5 CAPM and the present value of cash flows

Assume that we have a project j that will produce a cash flow in the next
period CF
g j,t+1 we want to evaluate the project and finding the correct price
that we should pay today pj,t . We know that

g j,t+1 − pj,t
CF
r̃j =
pj,t
 
E CF j,t+1 − pj,t
g
E (r̃j ) =
pj,t
 
E CF
g j,t+1
1 + E (r̃j ) =
pj,t
but from the CAPM we also know that

E (r̃j ) = rf + βj [E (r̃M ) − rf ]
1 + E (r̃j ) = 1 + rf + βj [E (r̃M ) − rf ]
We can now combine the two equations
 
E CF
g j,t+1
= 1 + rf + βj [E (r̃M ) − rf ] (1.2)
pj,t
 
E CF
g j,t+1
pj,t =
1 + rf + βj [E (r̃M ) − rf ]
And we see that the correct price of the company is the discounted value of
the expected cash flows using the equilibrium expected return in the discount
factor.

Exercise 1.1
(from the Exam of December 2007) There are two stocks A and B,
whose market prices are given by pA = 50$ and pB = 75$. Suppose returns
are described by the following single factor model without idiosyncratic risk
ri = r0 + βi (rm − r0 )
and consider the following parameters: µm = 10% per annum, r0 = 3% p.a.,
βA = 2, and βB = 1.5. What prices for the two stocks do you expect one year
from today? (4 points)
6 CHAPTER 1. GENERAL EQUILIBRIUM VALUATION.

Solution
The expected returns according to the model are

µj = r0 + βj (µm − r0 )

and substituting numbers

µA = 0.03 + 2 (0.07) = 17

µB = 0.03 + 1.5 (0.07) = 13.5

Therefore expected prices are

E (pj,t+1 ) = pj,t (1 + µj )

E (pA,t+1 ) = 50 (1.17) = 58.5

E (pB,t+1 ) = 75 (1.135) = 85.125

1.1.6 Risk Premium and diversification

The risk premium is proportional to the correlation between the asset and
the market portfolio return. This is the contribution of the asset to the risk
of the portfolio, its correlation with the other component of the portfolio.
And the specific risk?

µi = r0 + βi (E [e rM ] − r0 )
ri = r0 + βi (e rM − r0 ) + εi
σi = E (ri − µi )2
2
 

σi2 = E (r0 + βi (e rM ] − r0 ))2

 
rM − r0 ) + εi − r0 − βi (E [e
σi2 = E (βi (e rM ]) + εi )2
 
rM − E [e
σi2 = E βi2 (e rM ])2 + ε2i + 2εi βi (e
 
rM − E [e rM − E [erM ])
σi2 = βi2 E (e rM ])2 + E ε2i + 2βi E [εi (e
   
rM − E [e rM − E [e rM ])]
2 2 2 2
σi = βi σM + σε

For a portfolio

σp2 = βp2 σM
2 2
+ σεp
Where
1.2. THE ZERO-BETA CAPM (BLACK) 7

N
X
2
σεp = wi2 σεi
2

i=1

This term depends on the allocation, i.e. it is idiosyncractic, only the

systematic component has a remuneration.

1.2 The Zero-Beta CAPM (Black)

Does the CAPM still hold if we eliminate the assumption of the existence
of a risk free rate? Is general equilibrium a necessary condition to prove
the existence of the SML? Black (1972) has developed a valuation model for
risky assets that does not consider the existence of a risk free security and
proves the existence of a SML using only the first order condition implied by
Markowitz mean-variance portfolio problems. In order to demonstrate how
we can derive a pricing equation out of the mean-variance efficient frontier,
we state the following:

Proposition 1.2 For any frontier portfolio p, except the (absolute) mini-
mum variance portfolio, there exists (on the opposite arm of the frontier)
a unique frontier portfolio with which p has zero covariance. We will call
this portfolio the zero covariance portfolio relative to p and denote its
vector of portfolio weight as ZC (p).

Let’s now start the derivation by considering the standard definition of

covariance between two portfolios corresponding to the allocation vectors wp ,
wq :

Cov (r̃p , r̃q ) ≡ wp> V wq (1.3)

where V is the variance-covariance matrix.
Assume that p is a frontier portfolio, then the first order conditons for
the minimum variance frontier with n risky assets is:

wp = λV −1 µ + γV −1 1 (1.4)
where λ and γ are Lagrangian parameters are uniquely determined by the
constraints:

µ> wq = E (r̃q )
1> w q = 1
8 CHAPTER 1. GENERAL EQUILIBRIUM VALUATION.

Let’s substitute (1.3) in the (1.4)

 −1 >
Cov (r̃p , r̃q ) = λV µ + γV −1 1 V wq
= λµ> V −1 V wq + γ1> V −1 V wq
= λµ> wq + γ1> wq

so we can write

Cov (r̃p , r̃q ) = λE (r̃q ) + γ (1.5)

for p on the minimum variance and q is any portfolio efficient or inefficient.
Let’s now consider, as a specific portfolio q, the ZC (p)

Cov r̃p , r̃ZC(p) = λE r̃ZC(p) + γ = 0


γ = −λE r̃ZC(p)

Then from eq.(1.5) one can derive:



Cov (r̃p , r̃q ) = λ E (r̃q ) − E r̃ZC(p) (1.6)
Setting q = p one gets:

Cov (r̃p , r̃p ) = σp2 = λ E (r̃p ) − E r̃ZC(p)



(1.7)
and dividing (1.6) by (1.7) one gets



Cov (r̃p , r̃q ) λ E (r̃q ) − E r̃ZC(p)
= 

σp2 λ E (r̃p ) − E r̃ZC(p)


E (r̃q ) − E r̃ZC(p)
βpq =

E (r̃p ) − E r̃ZC(p)

Solving this equation for E (r̃q ) one gets

E (r̃q ) = E r̃ZC(p) + βqp E (r̃p ) − E r̃ZC(p)




An investor that considers an efficient portfolio p as a proxy for the market

portfolio M , will get a SML under the new set of assumptions that do not
require the existence of a risk free asset.



E (r̃q ) = E r̃ZC(p) + β E (r̃p ) − E r̃ZC(p)
1.3. BAB (FRAZZINI AND PEDERSEN, 2013) 9

where the beta can be computed as in the standard CAPM for any portfolio
or asset.

Cov (r̃p , r̃q )

βqp =
σp2
In other terms, the linearity of the SML is not an indication of an equilibrium
in the market, it is simply an implication of the efficiency of the portfolio p
selected as proxy of the market portfolio.
Starting from this observation, Roll established a famous critique, stating
that the existence of a CAPM general equilibrium is untestable in the light
of the following two observations:
• The linearity of the empirical SML is not sufficient to confirm the ex-
istence of the market equilibrium.
• The exact Market portfolio cannot be observed, in fact investor port-
folios includes assets that are illiquid or intangible like e.g. the human
capital, whose exact value cannot be exactly determined.

1.3 BAB (Frazzini and Pedersen, 2013)

The paper “Betting Against Beta” of Frazzini and Pedersen (2013)1 tries
and explain an important evidence of modern financial markets: because of
leverage constraints among investors, the behavior of tilting toward high-beta
assets makes risky high-beta assets require lower risk-adjusted returns than
low-beta assets, which would require leverage to be as profitable as the first.
The model considers two kinds of agents. The first group of them faces
leverage constraints and, therefore, overweights high-beta assets in order to
achieve the required return, making the price of these assets to increase, and,
as a consequence, their returns to decrease. Other agents can use leverage,
but face margin constraints. Unconstrained agents underweight (i.e. short-
sell) high-beta assets and buy low-beta assets that they can lever up. The
model thus implies a flatter security market line relative to the CAPM (Black,
Jensen, and Scholes, 1972), that is better explained by the CAPM with
restricted borrowing (as in Black (1972, 1993), Brennan (1971), and Mehrling
(2005)), where the slope depends on the tightness of the funding constraints
across agents.
The natural consequence of this evidence, in asset pricing effect terms, is
the possibility to exploit this friction considering the returns on a “betting-
against-beta” - BAB - factor. This factor is simply a portfolio that holds
1
This section is based on material extracted from the thesis of Andrea Occhipinti
10 CHAPTER 1. GENERAL EQUILIBRIUM VALUATION.

low-beta assets, leveraged to a beta of one, and shorts high-beta assets, de-
leveraged to a beta of one. Because of the above defined leverage constraints
faced by many investors, the model predicts that BAB factors have a positive
average return, increasing in the ex-ante tightness of constraints and in the
spread in betas between high- and low-beta assets.
Going more in details, Frazzini and Pedersen (2013) consider an overlapping-
generations (OLG) economy, with agents i = 1, ..., I born in each time period
t with wealth Wti , and living for two periods. Securities s = 1, ..., S are traded
by investors, in a way such that each security s pays dividends δts and has x∗s
shares outstanding. Each time period t, young agents then choose a portfolio
0
of shares x = (x1 , ..., xS ) , investing the rest of their wealth in a risk free
asset, with return rf . The utility maximization problem is then:

0 f γi 0
max x (Et (Pt+1 + δt+1 ) − (1 + r )Pt ) − x Ωt x (1.8)
2
where Pt is the vector of prices at time t, Ωt is the variance-covariance
matrix of (Pt+1 + δt+1 ), γ i denotes individual i’s risk aversion. Agent i is
subject to the portfolio constraint
X
mit xs Pts ≤ Wti (1.9)
s

requiring that some multiple mit of the total investment, i.e. of the sum
over s of the shares in each security multiplied by their respective prices,
must be less than the agent’s wealth. The investment constraint depends
then on the agent i: for example, some investors might not be allowed to use
leverage, i.e. might have mi = 1, others might not only not be allowed to use
leverage, but might also be obliged to have reserves in cash (so that mi > 1).
Assuming a competitive equilibrium where total demand equals supply.
i.e. i xi = x∗ , and deriving hence the FOC and solving for agent i as
P

0 = Et (Pt+1 + δt+1 ) − (1 + rf )Pt − γ i Ωxi − ψti Pt (1.10)

where ψ i is the Lagrangian multiplier of the portfolio constraint, we get
for each investor i that the corresponding optimal portfolio xi is given by

1 −1
xi = Ω (Et (Pt+1 + δt+1 ) − (1 + rf + ψti )Pt ) (1.11)
γi
Then, the equilibrium condition will imply that

1 −1
x∗ = Ω (Et (Pt+1 + δt+1 ) − (1 + rf + ψt )Pt ) (1.12)
γ
1.3. BAB (FRAZZINI AND PEDERSEN, 2013) 11

i
P
where the aggregate risk aversion γ is defined as 1/γ = i 1/γ and
ψt = i (γ/γ i )ψti is the weighted average Lagrange multiplier.
P
These mathematical results allow in the end to derive the equilibrium
price at time t as

Et (Pt+1 + δt+1 ) − γΩx∗

Pt = (1.13)
1 + rf + ψt
i
Recalling that the return of any security s is always defined as rt+1 =
i i
(Pt+1 +δt+1 )/Pti −1, s s M
that the expression for beta of a stock is βt = covt (rt+1 , rt+1 M
)/vart (rt+1 ),
M
and defining the return of the market index as rt+1 , it is possible to prove
the following crucial result (Frazzini and Pedersen, 2013):

Proposition I (High beta is low alpha)

(i) The equilibrium required return for any security s is

s
Et (rt+1 ) = rf + ψt + βts λt (1.14)
M
where the risk premium is λt = Et (rt+1 ) − rf − ψt and ψt is the average
Lagrange multiplier, measuring the tightness of funding constraints;

(ii) A security’s alpha with respect to the market is αts = ψt (1 − βts ). The
alpha decreases in the beta, βts ;

(iii) For an efficient portfolio, the Sharpe ratio is highest for an efficient
portfolio with a beta less than one and decreases in βts for higher betas
and increases for lower betas.

Explicitly, then, risk premia are affected by the tightness of agents’ port-
folio constraints, as measured by the average Lagrangian multiplier ψt . As
empirical evidence shows, then, tighter portfolio constraints, i.e. higher ψt ,
flatten the security market line by increasing the intercept and decreasing
the slope λt .
This theoretical framework allows then to introduce a factor, i.e. BAB,
that goes long low-beta assets and short high-beta ones. Before providing
the details on how to build such portfolio, the following theoretical result is
fundamental, as the basis of the “betting-against” theory:

Proposition II (Positive expected return of BAB)

The expected excess return of the self-financing BAB factor is positive
βH − βL
BAB
Et (rt+1 ) = t L H t ψt ≥ 0 (1.15)
βt βt
12 CHAPTER 1. GENERAL EQUILIBRIUM VALUATION.

and increasing in the ex-ante beta spread (βtH − βtL )/(βtL βtH ) and fund-
ing tightness ψt . βtH and βtL denote respectively the betas of the high-
and low-beta assets portfolios, with βtH > βtL .

Thus, a BAB portfolio earns a positive expected return on average: the

size of the expected return depends on the spread in the betas and how
binding the portfolio constraints are in the market, as captured by the average
of the Lagrange multipliers ψt .
In order to construct the BAB portfolio, following Frazzini and Pedersen
(2013), pre-ranking betas are computed with respect to the monthly returns
on the CRSP value-weighted market index, exploiting monthly return data
on stocks from the CRSP tape2 . Estimated betas for security i is given by
σ̂i
β̂tT S = ρ̂ (1.16)
σ̂m
where σ̂i and σ̂m are the estimated volatilities for the stock and the mar-
ket, and ρ̂ is their estimated correlation. Volatilities and correlation are esti-
mated separately for two main reasons, as in Frazzini and Pedersen (2013).
First, one-year rolling standard deviation for volatilities and five-year rolling
correlation are used, in order to account for the stylized fact of correlations
moving more slowly than volatilities3 . Second, one year of non-missing data
are required to estimate volatilities, while three years of non missing data
are sufficient to estimate correlations. This allows not to reduce the sample
too much, given the high number of observations of correlation otherwise
necessary to compute pre-ranking betas.
To reduce the presence of outliers, Vasicek (1973) and Elton, Gruber,
Brown, and Goetzmann (2003) are followed, shrinking the time series es-
timates of beta (β̂iT S ) towards the cross-sectional mean (β̂ XS ) through the
following

β̂i = wi β̂iT S + (1 − wi )β̂ XS (1.17)

where for simplicity wi = w = 0.6 and β XS = 1 for all periods and across
all assets. In any case, this shrinkage factor does not change the ranking
across stocks’ betas necessary to compute BAB.
The BAB factor is hence computed: this portfolio is long low-beta stocks
and short high-beta stocks. Stocks are indeed ranked in ascending order on
2
In the original paper daily data are actually used in order to build the BAB factor,
but because of computational problems related to the big amount of data, in this work
monthly data are used. The following explanation will thus be based on the use of monthly
data.
3
As, for example, De Santis and Gerard (1997).
1.3. BAB (FRAZZINI AND PEDERSEN, 2013) 13

the basis of their estimated betas. The ranked securities are assigned to
one of two portfolios, i.e. low-beta and high-beta, rebalanced every calendar
month. In each portfolio, stocks are weighted by the ranked betas (i.e. lower-
beta securities have larger weights in the low-beta portfolio and higher-beta
securities have larger weights in the high-beta portfolio). More in detail,
hence, let z denote the n × 1 vector of beta ranks zi = rank(βit ) in each
calendar month, and let z̄ = 10n z/n be the average rank, with n number
of securities in each period and 1n an n × 1 vector of ones. The portfolio
weights are thus given by a vector w such that, for the low-beta an high-beta
portfolios it holds that in each period

wH = k(z − z̄)+
(1.18)
wL = k(z − z̄)−
where k = 2/10n |z − z̄| is a normalizing constant and x+ and x− denote
the positive and negative elements of a vector x. Hence, by construction, in
each period it holds that 10n wH = 1 and 10n wL = 1. The BAB portfolio is
thus a self-financing portfolio that is long the low-beta portfolio and long the
high-beta portfolio, with returns given by

BAB 1 L 1 H
rt+1 = (r − rf ) − H (rt+1
L t+1
− rf ) (1.19)
βt βt
L
where rt+1 = r0 t+1 wL , rt+1
H 0
= rt+1 wH , βtL = βt0 wL and βtH = βt0 wH . The
expected return of this BAB portfolio is thus given by Proposition II above.
14 CHAPTER 1. GENERAL EQUILIBRIUM VALUATION.
Chapter 2

Partial Equilibrium Valuation

The Arbitrage Pricing Theory is a partial equilibrium approach based on two

assumptions

1. A linear factor model provides a satisfactory model stocks returns.

2. Arbitrageurs identify riskless speculation opportunities and trade until
they disappear. (Weak form of efficiency)

In this context a partial equilibrium model is developed without any refer-

ence to a single rational agent model maximizing a specific utility function.
On the contrary the model is based on a simple non-arbitrage condition:
asset prices will respect a certain relation because, otherwise, arbitrage pos-
sibilities would be available to investors. Or, to put it in a slightly different
way, investors exploit every possible arbitrage opportunity, and in doing this
they drive asset prices to a specific “efficiency level”.

2.1 The linear factor model

As far as the return generating process is concerned the framework can ac-
commodate both single-factor and multi-factor models

ri = bi0 + bi1 F1 + εi
ri = bi0 + bi1 F1 + bi2 F2 + · · · + bik Fk + εi

Where F1 is the return of the first risk factor in time t and bi1 is the sensitivity
(exposure) of security i to the first risk factor. The most general factor model
can be described as follows:The return generating process can be represented
as follows

15
16 CHAPTER 2. PARTIAL EQUILIBRIUM VALUATION

r = b0 +BF + ε
Where

     
r1 b10 b11 b12 ··· b1k
 r2   b20   b21 b22 ··· b2k 
r =  ..  b0 =  ..  B =  ..
     
.. ... .. 
 .   .   . . . 
rN bN 0 b N 1 bN 2 · · · bN k
   
F1 ε1
 F2   ε2 
F =  ..  ε =  .. 
   
 .   . 
FK εK

where the driving uncertainty satisfies the following hypotheses

E (εi εj ) = 0, i 6= j i, j = 1, ..., N ,
E ε2i = s2i < +∞, i = 1, ..., N ,

E (Fk εi ) = 0 k = 1, .., K i = 1, ..., N

It is important to note that the APT does not suggest any specific risk
factor for the return generating process. This makes the model very flexi-
ble but it also makes the model fragile and hard to assess empirically: the
selection of the relevant risk factors is arbitrary.
The vector of expected return is given by:

µ =E (r) = b0 + BE (F) (2.1)

while the securities’ variance-covariance matrix Σ is given by

Σ = BΩB0 + diag (v)

Ω = E (F−E[F]) (F−E[F])0
 

where Ω is the variance covariance matrix of the risk factors while diag (v)
denotes a diagonal matrix with diagonal elements determined by the vector
v = (s21 , ..., s2N ) with i-th elements equal to the residual variance s2i . In the
case of uncorrelated factors Fk , Cov (Fj , Fk ) = 0, hence Ω is a diagonal matrix
and the above expression of Σ can be simplified in an equivalent scalar (=non
vector) notation:
2.2. ARBITRAGE PRICING THEORY 17

K
X
Σii = V ar(ri ) = b2i,k V ar (Fk ) + s2i (2.2)
k=1
K
X
Σij = Cov (ri , rj ) = bi,k bj,k V ar (Fk ) (2.3)
k=1

Notice that any model with E (F) 6= 0, can be reduced to a model with new,
zero mean, factors F
e using the following change of variables:

e = F − E (F)
F
eb0 = b0 + BE (F)

without altering the estimation of the factor exposures B. In fact insertion

of the above change of variables shows that estimation of the two linear
regression models:

r= b0 + BF + ε
e 0 + B0 F
r= b e +ε

produces the same risk exposures B and B0 :

B0 = B

Hence it is not restrictive to limit the discussion to models with zero mean
factors.

2.2 Arbitrage pricing theory

We can define an arbitrage portfolio as a portfolio of assets which fulfills the
following properties :

Condition 2.1 (Zero Initial Investment) The sum of the values invested
in the n securities has to be equal to zero

n
X
wi = w> 1 =0
i=1

Condition 2.2 (No Systematic Risk) The weighted average sensitivity co-
efficient of the portfolio with respect to every risk factor has to be equal to
zero
18 CHAPTER 2. PARTIAL EQUILIBRIUM VALUATION

n
X
wi bij = 0 ∀j
i=1

or in vector notation:
w> |{z}
|{z} B = |{z}
0
1×N N ×K 1×K

Condition 2.3 (Arbitrage opportunity) a portfolio P fulfilling the above

conditions is said to be an (asymptotic) arbitrage portfolio if its expected gain
remains strictly positive as the number of securities in the portfolio increases
N → ∞:

w> 1= 0, w> B = 0,
= w> µ > 0
 (N ) 
lim E rp
N →+∞

where µ denotes the vector of expected gains.

The standard hypothesis of the APT model is that arbitrage opportuni-
ties are not present in the financial market. This is due to the equilibrium
observation that, in case it were present, arbitrageurs would try exploit the
free lunch thus generating a directional price pressure that gets rid of them.
Observe that the condition of absence of arbitrage opportunities is a weak
requirement of efficiency: it is implied by the existence of an equilibrium
(e.g. CAPM equilibrium is free from arbitrage opportunities) but it does not
imply the existence of a market equilibrium.
The absence of (statistical) arbitrage opportunities can be formulated as
follows:

Condition 2.4 (Zero Expected Gain) In a market without arbitrage op-

portunities a portfolio that does not require capital and does not yield risk
must have a zero expected gain.

w> µ = 0, w> B = 0 ⇒ µp = 0

It is important to remark that all the above equations are positively

homogeneous, i.e. do not change when are multiplied by a positive number.
While the conventional mathematical definition of arbitrage portfolio re-
quires that the portfolio is zero risk state of nature by state of nature, the
current formulation requires that an arbitrage portfolio has zero exposure to
systematic sources of risk.
2.2. ARBITRAGE PRICING THEORY 19

Note that the return of a portfolio composed by a finite number N of

securities can have zero initial investment and zero systematic risk but a
positive idiosyncractic (security specific) risk. In this case a positive expected
gain could be justified as a compensation for idiosyncractic risk. On the other
hand, these components are by definition uncorrelated among stocks hence,
as the number of securities in the portfolio increases and diverges N → +∞,
it is possible by the law of large numbers to select a linear combination of
securities such that the specific risk contribution converges to 0. This explains
why the no-arbitrage condition is required to hold in the asymptotic limit of
a diverging number of securities N → ∞.
The valuation by no arbitrage is a partial equilibrium approach since the
absence of arbitrage opportunities is a condition weaker that the existence
of an equilibrium in the market. With respect to the equilibrium solution
concept it has the advantage of requiring less hypotheses on the market struc-
ture and to be statistically verifiable. The price to be paid for this advantage
is the fact that an arbitrage based theory is less informative compared to
one based on a model of the market equilibrium. The main weknesses are
the following: first the theory is void of economic content since it does not
provide any guidance on the selection of priced factors; the second element of
weakness is the fact that arbitrage pricing theory produces only information
on the valuation of some assets relative to the valuation of a set of funda-
mental assets whose prices are observed and taken as given. Hence a more
traditional economic-based equilibrium approach is required to identify the
relevant pricing factors and to obtain information on the relation between
the asset price and its fundamental properties.

2.2.1 Statement and Proof of the APT

Suppose there are n stocks listed in the market and the number of relevant
risk factors is k. The main result to value stocks within the APT framework
proved by Ross is:

Theorem 2.1 If no asymptotic arbitrage opportunities exist in the market,

then as the number of assets n → ∞, the expected gain of asset i, µi , is a
linear combination of a constant component λ0 and a systematic component
obtained by multiplication of the sensitivity bij with the j − th factor risk
premium λj plus a non systematic component vi which converges to zero as
20 CHAPTER 2. PARTIAL EQUILIBRIUM VALUATION

n → +∞
K
X
µi = λ0 + bik λk +vi (2.4)
k=1
v
u
u1 X N
lim t v2 = 0
N →∞ N i=1 i

The second equation means that, as the number of securities in the market
diverges, N → +∞, the contribution of the term the contribution of the terms
νi to the risk-return tradeoff described by eq.(2.4) becomes negligible and the
vector relation linking expected returns to risk exposures B:

µ=λ0 1 + Bλ

becomes exact.

Proof. Let’s assume that we perform a cross-sectional regression of the

expected returns of the stocks on a vector of ones and the matrix of the
risk exposures, that is on the columns of the matrix [1, B] including also a
vector of ones 1 in order to estimate an intercept which is a component of
the expected return common to all the securities and independent from the
risk exposures. We are basically assuming that each expected gain can be
represented as follows
k
X
µi = λ0 + bij λj +vi
j=1

and we get OLS estimates for λ0 , λj , while the contribution vi is the residual
component of the risk premium.
Construct now a portfolio P that invests in stock i an amount propor-
tional to the residual of the regression νi . Specifically we will assume that the
P 1/2
N
allocation on asset i is given by wiv = √Nνkvki
, where kvkN = ν
i=1 i
2
is
N
a constant of proportionality introduced to normalize the portfolio weights
which measures the cross sectional standard deviation of the residuals.
We know that by construction in a OLS regression the error terms have
zero mean and are uncorrelated with the independent variables so we have
that
N n
X 1 X
wi = √ νi = 0
i=1
N kvk n i=1
2.2. ARBITRAGE PRICING THEORY 21

and
n N
X 1 X
wi bij = √ vi bij = 0 j = 1, ..K
i=1
N kvkN i=1
Hence from the portfolio perspective these two properties imply that the
portfolio P is characterized by zero initial investment and zero exposure to
each of the K systematic risk factors. These two properties indicate that the
portfolio P is a candidate to be an arbitrage portfolio.
Observe that the expected return of the portfolio P is given by:
N N K
!
X X X
E [rp ] = lim (wν )T · µ = lim wi µi = lim w i λ0 + bik λk +vi
N →∞ N →∞ N →∞
i=1 i=1 k=1

We can see that this is the sum of three components. The first one is related
to λ0 . Given the nature of our portfolio, this component is equal to zero, in
fact
N N N
X X νi λ0 X
lim wi λ0 = lim √ λ0 = lim √ νi = 0
N →∞
i=1
N →∞
i=1
kvkN N N →∞ N kvkN i=1

The second component of the expected gain is related to the exposure to the
systematic risk factors, and also in this case this component has a zero value
by construction:
N K N K K N
X X X νi X X λj X
lim wi bij λj = lim √ bij λj = lim √ bij νi = 0
n→∞
i=1 j=1
N →∞
i=1
kvkN N j=1 N →∞
j=1
kvkN N i=1

At this point we can rewrite the expected gain of our portfolio as

N
X
E [rp ] = µp = lim wi vi
N →∞
i=1
N
1 X
= lim √ vi2
N →∞ N kvkN i=1
s
kvk2N
PN
i=1 vi2
= lim √ = lim =0
N →∞ N kvkN N →∞ N

This implies that the theorem can be proved by contraddiction. In fact

assume that limiting expected return µp > 0, then this would imply that P
is an arbitrage portfolio and therefore we would have reached a contraddiction
22 CHAPTER 2. PARTIAL EQUILIBRIUM VALUATION

with the hypothesis that assumes absence of arbitrage opportunities. Hence

we must conclude that
s
PN 2
i=1 vi
µP = lim =0
N →∞ N
and the theorem is proved.

2.2.2 Economic Interpretation of the λ coefficients

In order to get an economic interpretation to these coefficients let’s assume
that in the economy we have two relevant risk factors K = 2), that the error
terms vi are equal to 0. For each security we have that

µi = λ0 + bi1 λ1 + bi2 λ2
Let’s now assume that we can build a portfolio p with a zero exposure to both
risk factors (bp1 = bp2 = 0). The expected gain (which equals the return) of
this portfolio would be µp = λ0 .The portfolio p is a portfolio with zero risk
(systematic or specific) and thus λ0 , the expected gain of this portfolio, by
no arbitrage, must be equal to the return of a risk-free investment in this
economy rf .
Let’s now build a different portfolio with an exposure equal to 1 to the first
factor and zero exposure to the second risk factor (bP1 ,1 = 1 and bP1 ,2 = 0).
This portfolio is named Factor Mimicking or (Replicating) Portfolio 1 . The
expected return of this portfolio is

µP1 = λ0 + λ1
µP 1 = rf + λ1
λ1 = µP1 − rf
hence λ1 is the excess return over the risk free of a portfolio with a unit
exposure to the first risk factor and zero exposure to all other factors. It can
thus be interpreted as the risk premium that the investor gets for assuming a
unit exposure to this particular risk factor. The same interpretation applies
to λ2 which will correspond to the exected compensation for an investment
with unt exposure to the second factor and zero exposure to the first one.

2.3 Traded portfolios as Risk factors

A method which turns to be important in the application consists in choosing
as factors excess returns of portfolios of stocks. Assume that the return
2.3. TRADED PORTFOLIOS AS RISK FACTORS 23

generating process in the market is

ri = bi0 + bi1 F1 + bi2 F2 + εi

F1 = rA − rf and F2 = rB − rf

where A and B are two traded portfolios. By taking the expectation of this
equation, one gets:

µi = bi0 + bi1 E (F1 ) + bi2 E (F2 )

= bi0 + bi1 (µA − rf ) + bi2 (µB − rf )

This expression must be compared with the APT expression of the expected
return:

µi = rf + bi1 λ1 + bi2 λ2
Both these equations must hold for any portfolio traded in the market and
in particular they must hold also for the riskless asset rf , A and B.
Consider first the riskfree security setting:

µi = rf , bf 1 = 0, bf 2 = 0

then, consistency with the previous two equations implies

bi0 = rf .

Similarly, if one selects

µi = µA , bA1 = 1, bA2 = 0 (2.5)

then the two equations imply

E (F1 ) = λ1 = µA − rf

and similarly by setting

µi = µB , bB1 = 0, bB2 = 1 (2.6)

one gets

E (F2 ) = λ2 = µB − rf .
The above equations are all implied by the condition that the factors have
been selected as excess return of traded portfolios and are particularly useful
24 CHAPTER 2. PARTIAL EQUILIBRIUM VALUATION

in order to simplify the empirical test of the factor model. Consider for
simplicity the test when there are only two uncorrelated relevant factors.
Then we test the equation
µi − rf = α + bi1 λ1 + bi2 λ2
and the hypothesis α = 0 is not statistically rejected for any traded security
i = 1, ..., N traded in the market.
As a second test, notice that the risk exposures bi1 ,bi2 can be computed
by running the time series regression:
ri,t − rf,t = bi1 F1,t + bi2 F2,t + εi,t .
In addition, since the risk exposures for the two portfolios A and B are known
and given by eqs.(2.5,2.6) the expected risk premium can be estimated by
considering the sample means:
T
T →+∞ 1X
λ1 = µA − rf ' (rA,t − rf,t )
T t=1
T
T →+∞ 1X
λ2 = µB − rf ' (rB,t − rf,t ) .
T t=1

2.4 Long-Short portfolios

A conventional method used to create portfolios correlated with a specific
firm characteristics (e.g. book to market price ratio) relies on the creation
of a portfolio that buys firms with a high level of the characteristics (high
book to market) and sells firms with a low level of the firm characteristics
(low book to market). In this way one gets a portfolio whose return is highly
correlated with the information conveyed by the firm characteristics used to
select the firms.
Consider first the single factor case K = 1. In this case, if the APT
equation is applied separately to the long and short position, one gets:
µLong
P − rf = bLong
P 1 λ1
µShort
P − rf = bShort
P 1 λ1

and taking the difference among these two equations:

 
µLong−Short = bLong
P1 − b Short
P1 λ1
2.4. LONG-SHORT PORTFOLIOS 25

and, under the natural condition that the risk exposure of the long port-
folio is positive, bLong
P1 > 0, while the risk exposure of the short portfolio
is negative, bShort
P1 < 0, one gets that the expected return of the long-short
portfolio
 is proportional
 to the risk premium λ1 with a proportionality factor
Long Short
bP 1 − bP 1 .
Consider now the case K > 1, where it is expected that there’s more
than a relevant firm characteristics priced by investors. In this case it is
desirable to create K portfolios, each one maximally correlated with a single
firm characteristics but well difersified with respect to all the other K − 1
factors.
The construction introduced by Fama and French to create these portfo-
lios goes as follows:

• For each characteristics, securities are ranked according to their risk

exposures and grouped toghether in portfolios classified by the quintile
(e.g. terzile, quintile or decile portfolios) with respect to each charac-
teristics. In the general case the number of portfolios created are

g × g × ...g = g K
| {z }
K times

Consider for example the K = 2 factor case. Call the first factor,
FA , and the second one factor B, FB . Then for each factor securities
are split in terzile groups g = 3, say High Medium and Low, for each
factor. Then one creates 9 portfolios corresponding to all possible pairs
of terzile portfolios, and the following 9 portfolios are determined:

(HA , HB ) (MA , HB ) (LA , HB )

(HA , MB ) (MA , MB ) (LA , MB )
(HA , LB ) (MA , LB ) (LA , LB )

• For each firm characteristics, it is possible to create a single factor mim-

icking portfolio out of the g K building blocks previously constructed.
A factor mimicking portfolio is by construction maximally correlated
with respect to a single factor and well diversified with respect to the
remaining ones. This portfolio is obtained as a long-short combination
of two portfolios. For example in the case K = 2

P ortFA = PALong − PAShort

P ortFB = PBLong − PBShort
26 CHAPTER 2. PARTIAL EQUILIBRIUM VALUATION

- The long position will be in a portfolio that equally weights all the
builiding block portfolios corresponding to the top quantile with
respect to the factor of interest. In the example the postions
corresponding to the long positions to replicate factors A and B
are
1
PAlong = [(HA , HB ) + (HA , MB ) + (HA , LB )]
3
1
PBlong = [(HA , HB ) + (MA , HB ) + (LA , HB )]
3
- The short position will be in a portfolio that equally weights all
the builiding block portfolios corresponding to the lowest quan-
tile with respect to the factor of interest. In the example one will
short-sell the two portfolios:
1
PAshort = [(LA , HB ) + (LA , MB ) + (LA , LB )]
3
1
PBshort = [(HA , LB ) + (MA , LB ) + (LA , LB )] .
3

2.5 APT & CAPM

Suppose that

1. The APT holds in a single factor environment

2. The risk factor is the excess return of the market portfolio over the risk
free rate: rm − rf

From the return generating process we have that

ri − rf = α + bi1 (rm − rf ) + εi
Since we assume that the market portfolio is the unique relevant factor
in the APT model α = 0

ri − rf = bi1 (rm − rf ) + εi
µi − rf = bi1 (µm − rf )
And we see that the CAPM equilibrium is implied in this model. Under
these conditions APT and CAPM are both valid models in this environment.
CAPM is more specific of the APT because it specifies also the number and
the expression of the factors which determine the expected returns.
Chapter 3

The pricing kernel

3.1 First order condition

Assume a finite state space Ω. The first order condition (F.O.C.) for a
concave differentiable utility maximizer is given by:
P
Et−1 [U 0 (Wt (a∗ )) (Rt − rf 1)] = 0

where W (a∗ ) is the optimal wealth. If we define a new varaible, called a

price deflator, as the normalized level of marginal utility:
U 0 (Wt (a∗ )) (ω)
mt (ω) =
Et−1
P
[U 0 (Wt (a∗ ))]
one gets that the F.O.C. implies:
P
Et−1 [mt (Rt − rf 1)] = 0

Notice also that by construction the price deflator must satisfy the con-
dition:
P
Et−1 [mt ] = 1

3.2 The risk neutral measure

If the process mt (ω) is restricted to the states where mt (ω) > 0, correspond-
ing to the states where the agent is non satiated, U 0 (Wt (a∗ )) > 0, then one
can define a new measure:
mt (ω) 1{mt (ω)>0}
Q (ω) =   P (ω)
EP mt (ω) 1{mt (ω)>0}

27
28 CHAPTER 3. THE PRICING KERNEL

that includes also the information on the marginal utility of the investor and
the valuation equation implied by the first order condition can be restated
as:
 
Q (ω)
P
Et−1 (Rt (ω) − rf 1) = (3.1)
P (ω)
X Q (ω)
P (ω) (Rt (ω) − rf 1) = 0
ω∈Ω
P (ω)
Q
Et−1 [Rt ] = rf 1

which states that under the measure Q (ω) all securities have an expected
rate or return equal to the risk free rate. For this reason the measure Q (ω) is
called the risk neutral measure. It is important to remark that the the use of
the risk neutral model is implied by the first order condition for a risk averse
agent. In other terms, the use of the risk neutral measure has nothing to do
with the assumption on the agent beahvior, that is typically characterized
by risk aversion. On the contrary, shifting from the measure P (ω) to the
measure Q (ω) produces the important advantage that the current price of
risky cash flows can be discounted with a riskless interest rate. In fact any
security issued at time t − 1 at price St−1 and liquidated at time t with a
dividend distribution of δt (ω) matures a stochastic rate of return between
time t − 1 and time t equal to:

δt (ω)
RS (ω) = −1
St−1

and by eq.(3.1) this implies:

1 + RS (ω) = (1 + rf )
Q
 
Et−1

hence:  
Q δt (ω)
Et−1 = (1 + rf )
St−1
which is equivalent to prove that
 
Q δt (ω)
St−1 = Et−1
(1 + rf )

the time t − 1 price St−1 of the stochastic dividend distributed at time t,

δt (ω), is equal to the risk neutral expectation of the dividend discounted
with the risk free rate of return.
3.3. APT AND RISK NEUTRAL VALUATION 29

3.3 APT and risk neutral valuation

Linear arbitrage pricing theory predicts the risk adjustment which has to
be applied in order to evaluate risky projects. Remarkably, also a linear
arbitrage pricing theory can be used to define a pricing kernel, in such a way
that linear beta pricing in the multifactor model and risk neutral valuation
give equivalent results. A linear beta pricing model is equivalent to:

E P [Rt − rf 1] = [B] [λ] (3.2)

| {z } |{z}|{z}
N ×1 N ×K K×1

where Rt = (R1,t , ..., RN,t )T is the vector of N traded securities, λ = (λ1 , ..., λK )T ,
B = [b1 , ..., bN ]T
B = E P (Rt − rf 1) (Ft − µ)0 V−1
 

where
V =E P (Ft − µ) (Ft − µ)0 .
 

Inserting the definition of the sensitivity matrix in 3.2, one gets:

(Rt − rf 1) (Ft − µ)0 V−1 [λ] = 0

 
P
Et−1 [Rt − rf 1] − Et−1
P

(Rt − rf 1) 1 − (Ft − µ)0 V−1 λ

P


Et−1 = 0

hence, if one defines a pricing kernel:

mAP T
:= 1 − (Ft − µ)0 V−1 λ


t (3.3)

then one can conclude that the APT equation can be restated as:
 AP T 
P
Et−1 mt (Rt − rf 1) = 0

and, assuming a finite state space Ω and mAP T (ω) > 0,∀ ω ∈ Ω one can
define a risk neutral measure:

Q (ω) = mAP
t
T
(ω) P (ω)

that prices securities in the same way as a multifactor AP T does. The

following one factor model example clarifies the relation between the two
approaches.
30 CHAPTER 3. THE PRICING KERNEL

3.4 CAPM and SDF

Consider a ”toy economy” where the number of states of nature is 2, ω =
up, down and P (up) = P (down) = 0.5 There are two traded securities in this
economy; a risk free security and a risky one. The risk free rate of return Rf
is equal to 2%. The payoff of a Traded Risky security T R is given by:

TR $
up 3
down 2

and the current price is 2.3$. A corporation is considering the opportunity

to run a New Business whose cash flows are expected to be:

NB $
up 2.8
down 2.4

and its expected return µN B = 4. 475 3 × 10−2 .

In light of the above information one can also compute the expected
return µT R and the volatility σT R for T R that are given by:

µT R = 8. 695 7 × 10−2

1/2
σT R = 4. 725 9 × 10−2 = 0.217 39

and the fair price of the opportunity N B is given by:

E P [S1 (ω)]
PN B = (3.4)
1 + µN B
0.5 × 2.8 + 0.5 × 2.4
PN B = = 2. 488 6
1 + 4. 475 3 × 10−2
Assuming that a CAPM equilibrium holds in the market and that T R is
the market portfolio, one can verify that N B is in the set of efficient oppor-
tunities verifying that it is perfectly correlated with the market portfolio, in
fact:

σN B = 8. 036 7 × 10−2
Cov (RN B (ω) , RT R (ω)) = 1. 747 1 × 10−2
1. 747 1 × 10−2
ρ N B ,T R = = 1.0
8. 036 7 × 10−2 × 0.217 39
3.4. CAPM AND SDF 31

Note that the new business is expected to belong to a risk class described by
a risk exposure βN B = 0.369 69 hence its expected return is determined by
the SML equation and is given by:

µN B = 4. 475 3 × 10−2 = 0.369 69 8. 695 7 × 10−2 − 0.02 + 0.02

Now it is possible to verify that the same valuation can be obtained

introducing a risk neutral measure using eq.(3.3). If one sets:

mCAP M (ω) = 1 − Ψ (RT R (ω) − µT R )

µT R − µf
Ψ =
σT2 R

then the corresponding risk neutral measure Q can be computed and is given
by:

m (up) = 1 − Ψ (RT R (up) − µT R )

 
3 0.5 × 2 + 0.5 × 3
= 1 − 1. 417 7 − = 0.691 8
2.3 2.3
Q (up) = m (up) P (up)
= 0.5 × 0.691 8 = 0.345 9
m (down) = 1 − Ψ (RT R (down) − µT R )
 
2 0.5 × 2 + 0.5 × 3
= 1 − 1. 417 7 − = 1. 308 2
2.3 2.3
Q (down) = m (down) P (down)
= 0.5 × 1. 308 2 = 0.654 1

and the verification that the fair price PN B is also obtained as the risk neutral
discounted expectation of the cash flows is given by:

E Q [S1 (ω)] 0.345 9 × 2.8 + 0.654 1 × 2.4

S0 = = = 2. 488 6
1 + Rf 1.02
which is equal to the value computed in (3.4) using the CAPM equation to
compute the the discount rate adjustment for risk.
It is possible to verify that the two formulas are equivalent, repeating
the argument used to determine (3.3) that is the following. Consider the
valuation equation:
E P [S1 (ω)]
S0 =
1 + µN B
32 CHAPTER 3. THE PRICING KERNEL

it can be rearranged as follows:

 
S1 P
E = (1 + µN B )
S0
E P 1 + RN B − β (µM − µf ) = 1 + µf
 

E P RN B − β (µM − µf ) = µf
 


 N B  Cov P RN B , RM
P
E R − (µM − µf ) = µf
V arP (RM )
 (µM − µf )
E P RN B − E P RN B (RM − µM )
  
= µf
V arP (RM )
E P (1 − Ψ (RM − µM )) RN B = µf
 

and, setting:
E P mCAP M RN B − µf


= 0
mCAP M = (1 − Ψ (RM − µM ))
the verification that the two valuation approaches are equivalent is concluded.

3.5 Physical and risk-neutral probabilities

In the pricing of derivative instruments and contingent claims we usually use
the concept of ”risk neutral probability”. We will now explain the economic
intiution behind the relationship that exists between physical and risk-neutral
probabilities. Consider the situation of a state space with 2 states of nature
The stochastic variable m (ω) is called ”pricing kernel ” and is defined, for
each state of the world k such that mk as
U 0 (Y1,up ) U 0 (Y1,down )
m (up) = , m (down) =
E [U 0 (Y1 )] E [U 0 (Y1 )]
For an investor with no satiation (positive U 0 (Y )) we have that mk > 0 In
the simplified scenario we have that:

E [m (1 + r)] = (1 + rf )
π u × mu × (1 + ru ) + π d × md × 1 + rd = (1 + rf )

where π d = 1 − π u and

U 0 Y1d


u U 0 (Y1u ) d
m = and m =
E [U 0 (Y1 )] E [U 0 (Y1 )]
3.5. PHYSICAL AND RISK-NEUTRAL PROBABILITIES 33

where E [U 0 (Y1 )] = π u × U 0 (Y1u ) + π d × U 0 Y1d . The constraint implies:

π u × mu +
π d × md = 1
π u × U 0 (Y1u ) π d × U 0 Y1d

+
= 1
π u × U 0 (Y1u ) + π d × U 0 Y1d π u × U 0 (Y1u ) + π d × U 0 Y1d

Let’s focus now on the economic interpretation of mk : this is the ratio

between the marginal utility of wealth in state of the world k and the average
marginal utility of wealth in all the state of the world. A large number would
show that for me wealth is particularly “valuable” in this state of the world.
Given the decreasing marginal utility of wealth implied in the concavity of
the utility function we know that in state of the world k we are relatively
poorer and any additional marginal amount of wealth has a significant impact
on our utility.
We can now define a new quantity qk = mk πk . We know that

2
X 2
X
qk = mk π k = 1
k=1 k=1
mk π k ≥ 0 ∀k = 1, 2

So we can interpret qk as a new probability measure of state of the world

k. A specificity of this probability measure is that when I calculate the
expected return for every risky asset I get the risk free rate:

π u × mu × (1 + ru ) + π d × md × 1 + rd


= (1 + rf )
q u × (1 + ru ) + q d × 1 + r d


= (1 + rf )

How is this possible? We can understand this apparent contradiction

remembering that:

• A risky asset can be seen as a vector of payoffs in the different states

of the world.

• We give more importance (and so we are willing to pay more) to secu-

rities that pay high payoffs in states of the world where the marginal
utility of wealth is higher.

• These securities will have higher prices and lower expected returns just
for this differential preference for wealth in different states of the world.
34 CHAPTER 3. THE PRICING KERNEL

• If we capture this information in the probability measure we automati-

cally eliminate it’s effect from the expected return: the expected return
is no longer influenced by our preferences for wealth in different states
of the world because this information is included in the probabilities
that we use to discount the different states of the world.
We could also observe that the quantity
mk πk
ADk = , k = 1, .., l
(1 + rf )
is the price (positive for an investor with U 0 > 0) that an investor is willing
to pay for a security that generates a marginal increase of wealth in state of
the world k and no change of wealth in others states of the world. These
assets are called “Pure Securities” or “Arrow - Debreu” Securities.

3.5.1 A numerical example

An investor (with logarithmic utility function) has a portfolio of real assets
(for example real estate properties) whose next period value has the following
distribution:

S Y1s
u 10000
d 7000
The two states of the world have the same probability. The investor can
buy one of the two listed risky assets (A and B) whose terminal values follow
the following distribution

S As1 B1s
u 1 0.5
d 0.5 1

Problem 3.1 How much is the investor willing to pay for the two assets?
We can see how the expected utility of the investor changes when the two
assets are added to the initial portfolio.
If we consider the initial risky situation the expected utility is:

1 1
ln (Y1u ) + ln Y1d


E [U (Y )] =
2 2
1 1
E [U (Y )] = ln (10000) + ln (7000) = 9.032003
2 2
3.5. PHYSICAL AND RISK-NEUTRAL PROBABILITIES 35

and the certain equivalent is

E [U (Y )] = U (CE)
9.157660 = ln (CE)
CE (Y ) = exp (9.032003) = 8366.600

Now let’s consider what happens if the investor buys the security A

1 1
E [U (Y + A)] = ln (10001) + ln (7000.5) = 9.032089
2 2
CE (Y + A) = exp (9.032089) = 8367.317

We can consider, as the marginal price of A the difference between the two
certain equivalents

PA = CE (Y + A) − CE (Y ) = 0.717
For security B we have that

1 1
E [U (Y + B)] = ln (10000.5) + ln (7001) = 9.032099
2 2
CE (Y + B) = exp (9.032099) = 8367.407
PB = CE (Y + B) − CE (Y ) = 0.807

The investor is willing to pay more for security B because the payoff distri-
bution reduces the volatility of the terminal wealth. Security B can be seen
as an insurance against the volatility of the initial portfolio.
Given these prices we can calculate the expected returns of the two risky
assets

Au1 − PA Ad − PA
E (rA ) = π u × + πd × 1
PA PA
1 1 − 0.717 1 0.5 − 0.717
= × + × = 4.583%
2 0.717 2 0.717

B1u − PB B d − PB
E (rB ) = π u × + πd × 1
PB PB
1 0.5 − 0.807 1 1 − 0.807
= × + × = −7.037%
2 0.807 2 0.807
36 CHAPTER 3. THE PRICING KERNEL

Remark: do not be surprised that the expected returns are, sometimes,

negative. This results stems from the choice of the logarithmic utility function
and even if they seem to be a bit unrealistic they still can be used in our
demonstration.

Problem 3.2 How can we estimate the risk neutral probabilities in this sys-
tem?

We know that

q u = π u × mu and q d = π d × md
0 0 d


u U Y
U (Y )
qu = πu × and q d = π d ×
E [U 0 (Y )] E [U 0 (Y )]

given the logarithmic utility function we have that

0 1
U (Y u ) = = 0.000100
10000
0 1
U Yd =


= 0.000143
7000
h 0 i 1 1
E U (Y ) = × 0.0001 + × 0.000143 = 0.000 121
2 2
0.0001 0.000143
mu = = 0.824 and md = = 1.176
0.000 121 0.000 121
Finally we have that

1
qu = × 0.824 = 0.412
2
1
qd = × 1.176 = 0.588
2
As you can see the new probabilities incorporate the different marginal
utility of wealth in the two states of the world. When we calculate average
returns or payoff under the new probability measure we give more importance
to what is paid in the state of the world d because in that scenario we are
poorer and we give a greater value to wealth.

Problem 3.3 What is the expected return of the two assets under the new
probability measure?
3.5. PHYSICAL AND RISK-NEUTRAL PROBABILITIES 37

Au1 − PA Ad − PA
E q (rA ) = q u × + qd × 1
PA PA
1 − 0.717 0.5 − 0.717
= 0.412 × + 0.588 × = −1.569%
0.717 0.717

B1u − PB B d − PB
E q (rB ) = q u × + qd × 1
PB PB
0.5 − 0.807 1 − 0.807
= 0.412 × + 0.588 × = −1.569%
0.807 0.807
We see that under the new probability measure the two assets have the
same return.

Problem 3.4 How can we interpret the return of the assets under the risk
neutral probability measure?

Let’s consider a risk free asset that produces a next period value of 0.75
in every possible state of nature. The price of this asset is

1 1
E [U (Y + RF )] = ln (10000.75) + ln (7000.75) = 9.032094
2 2
CE (Y + RF ) = exp (9.032094) = 8367.362
PRF = CE (Y + RF ) − CE (Y ) = 0.762

We can also calculate the expected return under the physical probabilities
RF − PRF 0.75 − 0.762
rf = = = −1.569%
PRF 0.762
So the return of the risky assets under the risk neutral probabilities is equal
to the return of a risk free asset.
38 CHAPTER 3. THE PRICING KERNEL
Chapter 4

Empirical tests

4.1 Empirical Test of Equilibrium Models

In the development of the equilibrium models previously analyzed we stressed
the fact that the construction of a theory necessitates a simplification of the
phenomena under study. While a model based on simple assumptions can
always be called into question because of these assumptions, the relevant test
of how much damage has been done by the simplification is to examine the
relationship between the predictions of the model and observed real world
phenomena. In our case, the relevant test is how well the simple CAPM or
APT describes the behavior of actual capital markets.
The problem now becomes how one design meaningful empirical tests of
a theory? In particular, how can one test the CAPM or any of its numerous
variants? In this chapter we review several of the tests of the equilibrium
models that have been presented in the literature. Finally, we will discuss
the fundamental critique moved by Roll (1977) to the general idea of testing
empirically an equilibrium model.

4.1.1 Ex-ante Expectations and Ex-post Tests

Most tests of general equilibrium models deal with either the standard CAPM
or the Zero Beta (two-factor) form of the CAPM.
The basic Sharpe-Lintner-Mossin CAPM can be written as

E (ri ) = r0 + βi [E (rM ) − r0 ]
The zero-beta version can be written as

E (ri ) = E (rZCM ) + βi [E (rM ) − E (rZCM )]

39
40 CHAPTER 4. EMPIRICAL TESTS

Where E (rZCM ) is the expected return on the minimum variance portfolio

that is uncorrelated with the market portfolio.
Notice that these models are formulated in terms of expectations. All
variables are expressed in terms of future values. The relevant Beta is the
future Beta on the security. Furthermore, both the return on the market and
the return on the minimum variance zero-correlation portfolio are expected
future returns.
Since large-scale systematic data on expectations do not exist, almost all
tests of the CAPM have been performed using ex-post or observed values for
the variables. This raises the logical question of how one justifies testing an
expectational model in terms of realizations.
There are two lines of defense that have commonly been used by re-
searchers. The simpler defense is to argue that expectations are on av-
erage and, on the whole, correct. Therefore, over long periods of time,
actual events can be taken as proxies for expectations.
The more complex defense starts by assuming that security returns are
linearly related to the return on a market portfolio. This model, called the
market model, can be written as

rit = αi + βi rM t + εit
and the expected return becomes

E (ri ) = αi + βi E (rM )
αi = E (ri ) − βi E (rM )
we can now substitute this result in the previous equation and we get

rit = E (ri ) + βi [rM t − E (rM )] + εit

Now from the standard form of the CAPM we now that

E (ri ) = r0 + βi [E (rM ) − r0 ]
Substituting the expected return in the previous equation and simplifying
yields

rit = r0 + βi [E (rM ) − r0 ] + βi [rM t − E (rM )] + εit

rit = r0 + βi [rM t − r0 ] + εit
So we can test the CAPM using time series of (historical) realized returns.
However the reader should notice that there are three assumptions behind
this model:
4.1. EMPIRICAL TEST OF EQUILIBRIUM MODELS 41

1. The return generating process (the market model) holds in every period

2. The CAPM holds in every period

3. The Beta is stable over time

When we test the CAPM using historical data we are actually testing
simultaneously all three hypothesis.

4.1.2 Some Hypotheses of the CAPM

Regardless of the specific form of the CAPM that we want to test there are
three main hypotheses that should hold

• The first is that higher risk (Beta) should be associated with a higher
level of return.

• The second is that return is linearly related to Beta; that is, for every
unit increase in Beta, there is the same increase in return.

• The third is that there should be no added return for bearing non-
market risk.

In addition, if some form of equilibrium model holds, then investing

should constitute a fair game with respect to it. That is, deviations of a
security or portfolio from equilibrium should be purely random and there
should be no way to use these deviations to earn an excess profit.
In addition to the hypotheses common to both the standard and the
zero-beta form of the CAPM, we can formulate hypotheses that attempt to
differentiate between these general equilibrium models. In particular, the
standard version implies that the security market line, drawn in a return-
beta space, should have an intercept of r0 and a slope of (r̄M − r0 ) , while
the two-factor version requires that it should have an intercept of r̄ZCM and
a slope of (r̄M − r̄ZCM ).

4.1.3 Common methodologies to test the CAPM

There are basically two commonly used methodologies to empirically test an
equilibrium model. The first one is based on time-series analysis while the
second utilizes cross-sectional data.
42 CHAPTER 4. EMPIRICAL TESTS

Time-series tests
In order to apply the first methodology we have to run a time-series regression
of the return of one asset (stock or portfolio) over the return of a proxy for
the market portfolio. There most commonly used specification of this test is
the following

rit − r0t = αi + βi [rM t − r0t ] + εit

The reader should note that now the risk free variable as time subscript.
This is because we do not have a real risk-free asset in the economy and we
usually proxy the risk free rate with the yield to maturity of a short term
government bond. This rate has a positive, albeit small, volatility. The
obvious testable implication of this model is H0 : α̂i = 0. That is there
should not be a positive return above the the one justified by the exposure
to the market risk factor.

Cross-sectional tests
The second widely used testing methodology is based on the regression of
a cross-section of equity returns on a proxy for the exposure to the market
risk factor and, possibly, additional risk factors. There are many possible
specification of this kind of tests, for example we could estimate the following
model

rit = b0t + b1t βi + b2t βi2 + b3t δi + εit

Where δi is the exposure of security i to another risk factor. The reader
should note that now the sample contain the returns of many stocks in a
single time period and thus the estimated coefficients have a time subscript
because they are not relative to a specific security but to a specific period
of time. There are many testable implication in this simple specification,
specifically:

• b̂0t = r0t or b̂0t = rZCM t , the specific hypothesis depends on which form
of the CAPM we are testing.

• b̂1t = (rM t − r0t ) or b̂1t = (rM t − rZCM t ), again as in the previous line
this test can help us to distinguish between the two forms of CAPM.

• b̂2t = 0, independently from the specific CAPM form the relationship

between market risk exposure and return should be linear.

• b̂3t = 0, other risk factors should not be priced.

4.1. EMPIRICAL TEST OF EQUILIBRIUM MODELS 43

As we can see the cross-sectional methodology is much more flexible than

the time-series approach and offers a variety of testable implication. As a
drawback this methodology requires the previous knowledge of the sensitivity
of assets returns to the risk factors but, as we know, the market beta of a
stock cannot be observed and must be estimated via a time series regression.
Cross-sectional tests acquire, thus, the nature of two-stage experiments: in
the first stage time series regressions are performed in order to estimate the
betas that will be used in the second stage cross-sectional regression.

4.1.4 An example of time-series test: Black, Jensen

and Scholes (1972)
Black, Jensen, and Scholes (1972) were the first to conduct an in-depth time
series test of the CAPM. They considered the following time series model

rit − r0t = αi + βi [rM t − r0t ] + εit

When this equation is estimated on time series data, the regression coef-
ficient αi should be equal to zero if the simple CAPM describes returns.
In order to test the CAPM, it is desirable to use a large number of secu-
rities. The obvious method is to estimate the equation for each of a series of
securities and then examine the distribution of αi However, this is inappro-
priate because tests of the distribution of αs assume that the residuals (εit ,
εjt ) are independent, and they are not.
One way to alleviate the problem is to run the time series regression
on portfolios. Now rit is the return on portfolio i. Since portfolios utilize
data on more than one security and, since the residual variance from the
regression using portfolios will incorporate the effect of any cross-sectional
interdependencies, the standard error of the intercept can be used to test the
difference of αi from zero.
When they form portfolios, Black, Jensen, and Scholes want to maximize
the spread in Betas across portfolios so they can examine the effect of Beta
on return. The most obvious way to do this is to rank stocks into portfolios
by true Beta. But all we have is observed Beta. To rank into portfolios by
observed Beta would introduce selection bias. Stocks with high observed Beta
(in the highest group) would be more likely to have a positive measurement
error in estimating Beta. This would introduce a positive bias into the Beta
for high Beta portfolios and would introduce a negative bias into an estimate
of the intercept αi In an attempt to avoid this problem, an instrumental
variable was used to rank stocks into portfolios. An instrumental variable is
one that ideally is highly correlated with the true Beta but can be observed
44 CHAPTER 4. EMPIRICAL TESTS

independently. The instrumental variable used in this study and, indeed in

most studies of the CAPM, is the Beta for each security in the previous time
period.
The exact procedure Black, Jensen, and Scholes is the following:

1. Five years of monthly data are used to estimate Betas and rank stocks
into deciles (from highest to lowest). Each decile is then considered as
one of the portfolios in the next (e. g., sixth) year.

2. The procedure was then repeated moving one year forward (data for
the second through sixth year were used to estimate betas rank and
form deciles that were considered as portfolios for the seventh year).
This was repeated for the entire length of the sample, 35 years.

3. Then the return for decile one in each year was considered a time-
series of returns from a portfolio, the return for decile two in each year
considered a series of returns on another portfolio, and so forth.

4. Each of the ten portfolios could then be regressed against the market
and an intercept, a Beta, and a correlation coefficient for the equation
computed.

The following table shows the excess return estimated intercept and slope
coefficients, and R2 for each decile reported by Black, Jensen, and Scholes.

Decile Portfolio Excess Return βi Alpha R2

1 0.0213 1.561 −0.0829 0.927
2 0.0177 1.384 −0.1938 0.976
3 0.0171 1.248 −0.0649 0.976
4 0.0163 1.163 −0.0167 0.982
5 0.0145 1.057 −0.0543 0.984
6 0.0137 0.923 0.0593 0.966
7 0.0126 0.853 0.0462 0.970
8 0.0115 0.753 0.0812 0.958
9 0.0109 0.629 0.1968 0.913
10 0.0091 0.490 0.2012 0.806
Market 0.0142 1.000

Note, first, how well the model explains excess returns (the high value of
the R2 s). This would tend to support the structure of the linear equation as a
good explanation of security returns. Note, however, that the intercepts vary
4.1. EMPIRICAL TEST OF EQUILIBRIUM MODELS 45

quite a bit from zero. In fact, when β > 1 the intercepts tend to be negative
and when β < 1 the intercepts tend to be positive. This, as explained below,
is more consistent with the zero-beta capital asset pricing model rather than
the standard CAPM.
The implications of the zero Beta form of the CAPM are that

rit − rZCM t = βi [rM t − rZCM t ] + εit

The model tested is

rit − r0t = αi + βi [rM t − r0t ] + εit

If the zero-beta model really explains security prices, then we can
rearrange the two equations in the following way

r̄i = r̄ZCM + βi [r̄M − r̄ZCM ]

r̄i = r̄0 + αi + βi [r̄M − r̄0 ]

we can now take the expected value of the two models, combine the two
right-hand sides and solve for αi

r̄0 + αi + βi [r̄M − r̄0 ] = r̄ZCM + βi [r̄M − r̄ZCM ]

αi = (1 − βi ) (r̄ZCM − r̄0 )

From our previous analysis of the CAPM we know that the zero-beta
portfolio has zero correlation with the market portfolio but has a positive
risk so we should expect (r̄ZCM − r̄0 ) > 0. Therefore, if β < 1, αi should
be positive; and if β > 1, αi should be negative. This is exactly what the
empirical results show. Black, Jensen, and Scholes repeat these tests for
four sub-periods and find, by and large, the same type of behavior we have
described for the overall period.

4.1.5 Fama and MacBeth (1973) cross-sectional ap-

proach
Fama and MacBeth (1973) introduced an interesting methodology to test the
CAPM that is now applied also to the analysis of the multifactor arbitrage
pricing theories. The basic methodological innovation relies on the observa-
tion that in an efficient market, cross-sectional correlations are expected to
be stronger than time series ones, which are reduced by traders that try to
46 CHAPTER 4. EMPIRICAL TESTS

exploit predictability patterns from the return time series in order to exploit
profitable investment opportunities.
The method considers the panel of returns Ri,t observed for different
securities and different times define a panel of data. Linear regressions must
take into account the potential biases arising from the cross sectional and the
temporal correlations. The Fama McBeth two pass strategy is very simple:

• First step: run a time series of cross sectional regressions:

• Second step: compute sample means of regression coefficients and sam-

ple standard regressions to measure the significance of the coefficients.

In vector notation the method works as follows:

• Assume that the vector of excess expected returns is given by µbt −

bf I ∈RN . The risk exposures B
µ b are estimated using the a window of
S observations between time t − 1 and time t − S, running the cross
sectional estimation:
bt − µ
µ bf I =Bγ
b + et

• Then:   −1   
0 0
bt = B
γ b B b B
b (b µt − µ
bf I)

and the final estimate is determined by the time series mean of the
estimators:
T T
1 X 1 X
γ
b= γ
bt e= et
T − S t=S+1 T − S t=S+1
b b

4.1.6 A first example of cross-sectional test

They formed 20 portfolios of securities to estimate Betas from a first-pass
regression. They then performed one second-pass cross-sectional regression
for each month subsequent to the estimation period over the time period
1935-1968. The equation they tested was

rit = γ0t + γ1t βi + γ2t βi2 + γ3t Sei + ηit

where Sei is the standard deviation of the residuals of the first-pass time-
series regression for security i. By estimating this equation (in cross section)
for each month, it is possible to study how the parameters change over time.
This form of the equation allows the test of a series of hypotheses regard-
ing the CAPM. The tests are:
4.1. EMPIRICAL TEST OF EQUILIBRIUM MODELS 47

1. E (γ̂3t ) = 0, residual risk does not affect return.

2. E (γ̂2t ) = 0, there are no nonlinearities in the security market line.

3. E (γ̂1t ) > 0, there is a positive price of risk in the capital markets.

If both E (γ̂2t ) and E (γ̂3t ) are not different from zero, we can also examine
both E (γ̂0t ) and E (γ̂1t ) to see whether the standard CAPM or zero-beta
model is a better description of market returns.
Finally, we can examine all of the coefficients and the residual term to
see if the market operates as a fair game. If the market is a fair game, then
there is no way that one should be able to use knowledge about the value of
the parameters in previous periods to make an excess return. For example,
if the CAPM holds, then, regardless of the prior values of γ̂2t and γ̂3t each of
their expected values at time t + 1 should be zero. Furthermore, if the zero-
beta model is the best description of general equilibrium, then deviations of
γ̂0t from its mean E (rZCM ) and γ̂1t from its mean E (rM ) − E (rZCM ) are
random, regardless of what happened at time period t − 1 or any earlier time
period. If the simple form of the CAPM holds, the same statements should
be true with r0 substituted for rZCM .
Fama and MacBeth have estimates of all the coefficients for each month
over the period January 1935-June 1968. The average value of any γ̂it (de-
noted by γ̄it ) can be found simply by averaging the individual values, and
this mean can be tested to see if it is different from zero.
Fama and MacBeth (1973) presents the results of estimating the previous
equation and several variations of it over the full time period as well as for
several sub-periods. Here we will report only results for the full model and
the entire estimation period, the reader should check the original paper fore
more details.

Coefficient Mean Value t-stat Serial Correlation

γ̂3t 0.0516 1.11 −0.12
γ̂2t −0.0026 −0.86 −0.09
γ̂1t 0.0085 2.57 0.02
γ̂0t − r0t 0.0048 2.55 0.15

Empirical results reveal that, when measured over the entire period, γ̄3t
is small and is not statistically different from zero. Furthermore, when we
examined over several sub-periods, we find that it remains small in each sub-
period, is not significantly different from zero, and, in fact, exhibits different
signs in different sub-periods. We can safely conclude that residual risk has
no effect on the expected return of a security. However, it is still possible that
48 CHAPTER 4. EMPIRICAL TESTS

the market does not constitute a fair game with respect to any information
contained in γ̂3t That is, it is possible that the fact that γ̂3t differs from zero
in any period gives us insight into what its value (and, therefore, returns)
will be next period. The easiest way to test this is to examine the correlation
of γ̂3t in one period with its value in the prior period where the mean of all
periods is assumed to be zero. Again the analysis shows that the value of
this correlation coefficient is close to zero and not statistically significant.
Fama and MacBeth also compute the correlation between γ̂3t and its prior
value for lags of more than one period. They find, once again, that there is
no usable information contained in γ̂3t .
The results, with respect to γ̄2t , are very similar. The average coefficient
is small, is not statistically significant, and changes sign over alternative
sub-periods. Furthermore, an examination of the correlation of γ̂2t with
its previous value (with means assumed to be zero) shows that there is no
information contained in individual values of γ̂2t Thus, the Beta squared
term neither affects the expected return on securities, nor does its coefficient
contain information with respect to an investment strategy.
Having concluded that neither Beta squared nor residual risk has an in-
fluence on returns, the correct form of the equation to examine for further
tests is the following:

rit = γ0t + γ1t βi + ηit

Fama and MacBeth estimate this reduced model and examine the perfor-
mance of γ̂1t for the entire period and conclude that there is evidence that the
relationship between expected return and beta is positive as well as linear.
Furthermore, by testing the correlation of the difference between γ̂1t and its
mean with prior values of the same variable, they show that difference in γ̂1t
from its mean cannot be employed to produce a better forecast of a future
value of γ̂1t than simply using the mean.
Fama and MacBeth find that γ̂0t is generally, and significantly, greater
than a reasonable proxy of the risk free rate (r0t ) and in addition, they find
that γ̂1t is generally less than rM t − r0t . The fact that γ̂0t is substantially
greater than r0t and γ̂1t is substantially less than rM t −r0t would seem to indi-
cate that the zero-beta model is more consistent with equilibrium conditions
than is the simple CAPM.
Before finishing our discussion of these tests one more point is worth
mentioning. If the equilibrium model describes market conditions, then an
individual security’s deviation from the model should contain no information.
That is, a positive residual value for any one stock at any moment in time
should convey no information about the differential performance of that stock
4.2. FAMA AND FRENCH (1992) 49

(from the expected value produced by the model) in future periods. For this
to be true, there should be no correlation (with any lag) between the residuals
in the previous equation. This is, in fact, what Fama and MacBeth found.

4.1.7 The Roll’s Critique

Roll (1977) moves an important critique to the possibility of empirically test-
ing the validity of the CAPM. The intuition of this critique can be addressed
in two points.
The first one can be understood going back to the derivation of the Black
CAPM. In that case we demonstrated that a linear pricing relationship could
be derived starting from any mean-variance efficient portfolio p.



E (ri ) = E rzcp + βi E (rp ) − E rzcp
In other words mean-variance efficiency implies a the existence of a linear
pricing relationship but for this relationship to be the CAPM it must be
derived starting from the market portfolio.
The second point of Roll’s critique is precisely the fact that the real
market portfolio is unobservable. From statement one, validity of the CAPM
is equivalent to the market being mean-variance efficient with respect to all
investment opportunities. Without observing all investment opportunities,
it is not possible to test whether this portfolio, or indeed any portfolio, is
mean-variance efficient. Consequently, it is not possible to test the CAPM.
A less caustic phrasing of the Roll’s critique is that actually every empir-
ical test of the CAPM is actually a joint test on the pricing model AND on
the choice of the proxy for the market portfolios.

4.2 Fama and French (1992)

More recently Eugene Fama and Kenneth French analyzed the cross-section
of equity returns in order to test the significance of risk factors in addition
to the market portfolio. The starting point of their work is the consideration
of the two main predictions of the CAPM:

• Expected returns on securities are a positive linear function of their

market βs (the slope in the regression of a security’s return on the
market’s return).

• Market βs suffice to describe the cross-section of expected returns.

50 CHAPTER 4. EMPIRICAL TESTS

They then consider that there are several empirical contradictions of the
Sharpe-Lintner-Black (SLB) model. The most prominent is the size effect
of Banz (1981). He finds that market equity, ME (a stock’s price times shares
outstanding), adds to the explanation of the cross-section of average returns
provided by market βs. Average returns on small (low capitalization) stocks
are too high given their β estimates, and average returns on large stocks are
too low. Another contradiction of the CAPM is the positive relation between
leverage and average return documented by Bhandari (1988). It is plausible
that leverage is associated with risk and expected return, but in the CAPM
leverage risk should be captured by market β. Bhandari finds, however, that
leverage helps explain the cross-section of average stock returns in tests that
include size as well as β. Finally Stattman (1980) and Rosenberg, Reid,
and Lanstein (1985) find that average returns on U.S. stocks are positively
related to the ratio of a firm’s book value of common equity, BE, to its
market value, ME. Chan, Hamao, and Lakonishok (1991) find that book-
to-market equity, BE/ME, also has a strong role in explaining the cross-
section of average returns on Japanese stocks. Finally, Basu (1983) shows
that earnings-price ratios (E/P) help explain the cross-section of average
returns on U.S. stocks in tests that also include size and market β.
Fama and French compile all the previous research in a single cross-
sectional experiment using data on all non-financial firms listed on the NYSE,
AMEX, and NASDAQ from 1962 to 1989.
They replicate on this new sample the cross-sectional regression approach
of Fama and MacBeth (1973): each month the cross-section of returns on
stocks is regressed on variables hypothesized to explain expected returns. The
time-series means of the monthly regression coefficients then provide standard
tests of whether different explanatory variables are on average priced. Since
size, E/P, leverage, and BE/ME are measured precisely for individual stocks,
the experiment is at the equity level instead of using portfolios as in previous
research. Most previous tests used portfolios because estimates of market βs
are more precise for portfolios. The Fama and French approach is to estimate
βs for portfolios (using 5 years of past monthly observations) and then assign
a portfolio’s β to each stock in the portfolio.
The author perform a significant amount of empirical univariate and mul-
tivariate analysis in order to test the explanatory power of the different risk
factors. They conclude that the most relevant factor are the (natural loga-
rithm of) equity market capitalization (ME) and the (natural logarithm of)
book-to-market ratio (BE/ME).
These two characteristics are termed CAPM anomalies in fact it is pos-
sible to use this information to sort securities in such a way that: portofolio
determined by the sort (e.g. quintile or terzile portfolio) have increasing
4.2. FAMA AND FRENCH (1992) 51

expected excess returns and ii) portfolio sorts generate a siginficant alpha
or equivalently risk exposure, the CAPM beta, cannot explain these excess
returns.
We can summarize the main results of the estimation of the following
cross-sectional model
 
BE
rit = a + b1t βi + b2t ln (M E) + b3t ln + eit
ME

Jul 63 - Dec 90 Jul 63 - Dec 76 Jan 77 - Dec 90

Variable Mean t-Stat Mean t-Stat Mean t-Stat
a 2.07 6.55 1.73 3.54 2.4 5.92
b1 −0.17 −062 0.10 0.25 −0.44 −1.17
b2 −0.12 −2.52 −0.15 −1.91 −0.09 −1.64
b3 0.33 4.8 0.34 3.17 0.31 3.67

As we can see the author demonstrate that size and book-to-market ratio
are stronger factors than the exposure to the market risk in the explanation
of the cross-section of equity returns.
Since this seminal work many authors have applied the same framework
to different samples showing that the relevance of these two additional risk
factors is not limited to the US market.

Beyond Fama and French (1992)

The existence of multiple risk factors priced in the stock market immediately
recall the idea of an APT-like equilibrium. As we have seen the testable
implications of the APT become much more powerful when the risk factors
are stock portfolios. The original F&F model considered two additional risk
factors not represented by equity returns. In order to increase the power
of their model the same authors developed a slightly different version of
their model where all the risk factors are represented by the return of equity
portfolios. The originating return generating process is

ri,t − rf,t = αi + βi,m (rm,t − r0t ) + βi,smb SM B + βi,hml HM L + εit

SM B = (rs,t − rb,t )
HM L = (rh,t − rl,t )

where:

• rm,t is the return of a proxy for the market portfolio

52 CHAPTER 4. EMPIRICAL TESTS

• rs,t is the return of a portfolio of stocks with low market capitalization

(small caps)

• rb,t is the return of a portfolio of stocks with high market capitalization

(large, or big, caps)

• rh,t is the return of a portfolio of stocks with high book-to-market ratio

(value stocks)

• rl,t is the return of a portfolio of stocks with low book-to-market ratio

(growth stocks)

In order to create the SMB and HML portfolios, Fama and French sorted
securities in three terzile groups w.r.t. the size factor (i.e. sorting varaible
ln (M E)) and two groups w.r.t. the value factor (i.e. sorting variable
ln (BE/M E)) creating six portfolios:

(SSM B , HHM L ) (MSM B , HHM L ) (BSM B , HHM L )

(SSM B , LHM L ) (MSM B , LHM L ) (BSM B , LHM L )

The SMB factor is a long-short portfolio such that:

1 
rs = r(SSM B ,HHM L ) + r(SSM B ,LHM L )
2
1 
rb = r(BSM B ,HHM L ) + r(BSM B ,LHM L )
2
being the difference between the performance of small and large caps, cap-
tures the idea of a risk premium based on size, while the HML factor captures
the idea of a risk premium based on the book-to-market ratio and its return
is given by:
1 
rh = r(SSM B ,HHM L ) + r(MSM B ,HHM L ) + r(BSM B ,HHM L )
3
1 
rl = r(SSM B ,LHM L ) + r(MSM B ,LHM L ) + r(BSM B ,LHM L )
3
These two factors, being defined as differential returns, should be un-
correlated with the market factor.
In 1997 Carhart suggested an augmented version of this three factor model
considering, as a forth factor, the difference between ru , the return of a
portfolio of stocks with a very good past performance and rd , the return of
a portfolio of stocks with a very poor past performance. This factor, usually
called U M D (up minus down) captures the existence of a momentum effect
in the market.
4.2. FAMA AND FRENCH (1992) 53

In 2015 Fama and French proposed a new model a 5 factor model. IN

addition to the traditional three factors, two new factors have been included,
CM A (Conservative Investment minus Agressive investment) and RM W
(Robust minus Weak operating profitability).
The Home page of Kenneth R. French contains the updated time series
which are necessary to repeat all the tests which make use of the ”Fama
French three factor model”.

ESSENTIAL BIBLIOGRAPHY

1. Black, F., M. Jensen and M. Scholes, 1972, The Capital Asset Pric-
ing Model: Some Empirical Tests, in Studies in the Theory of Capital
Markets, Michael C. Jensen, ed. New York: Praeger, 79-121.

2. Carhart, M., 1997, On Persistence in Mutual Fund Performance, Jour-

nal of Finance, 52, 57-82.

3. Fama, E., and J. MacBeth, 1973, Risk, return and equilibrium: Em-
pirical tests, Journal of Political Economy, 81, 607 -636.

4. Fama, E. and K. French, 1992, The Cross-Section of Expected Stock

Returns, Journal of Finance, 47, 427-465.

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

6. Roll, R. e Ross, S. (1980) An Empirical Investigation of the Arbitrage

Pricing Theory, Journal of Finance 35(4), 1073-1103.

7. Ross, S. (1976) The arbitrage theory of capital pricing, Journal of Eco-

nomic Theory 13.

8. Frazzini, Andrea, and Lasse Heje Pedersen. ”Betting against beta.”

Journal of Financial Economics 111.1 (2014): 1-25.