IEOR E4706: Foundations of Financial Engineering

c 2016 by Martin Haugh

Mean-Variance Optimization and the CAPM

These lecture notes provide an introduction to mean-variance analysis and the capital asset pricing model
(CAPM). We begin with the mean-variance analysis of Markowitz (1952) when there is no risk-free asset and
then move on to the case where there is a risk-free asset available. We also discuss the difficulties of
implementing mean-variance analysis in practice and outline some approaches for resolving these difficulties.
Because optimal asset allocations are typically very sensitive to estimates of expected returns and covariances,
these approaches typically involve superior or more robust parameter estimation methods.
Mean-variance analysis leads directly to the capital asset pricing model or CAPM. The CAPM is a one-period
equilibrium model that provides many important insights to the problem of asset pricing. The language / jargon
associated with the CAPM has become ubiquitous in finance.

1 Markowitz’s Mean-Variance Analysis

Consider a one-period market with n securities which have identical expected returns and variances, i.e.
E[Ri ] = µ and Var(Ri ) = σ 2 for i = 1, . . . , n. We also suppose Cov(Ri , Rj ) = 0 for all iP
6= j. Let wi denote
n
the fraction of wealth invested in the ith security at time t = 0. Note that we must have i=1 wi = 1 for any
portfolio. Consider now two portfolios:
Portfolio A: 100% invested in security # 1 so that w1 = 1 and wi = 0 for i = 2, . . . , n.
Portfolio B: An equi-weighted portfolio so that wi = 1/n for i = 1, . . . , n.
Let RA and RB denote the random returns of portfolios A and B, respectively. We immediately have
E[RA ] = E[RB ] = µ
Var(RA ) = σ2
Var(RB ) = σ 2 /n.
The two portfolios therefore have the same expected return but very different return variances. A risk-averse
investor should clearly prefer portfolio B because this portfolio benefits from diversification without sacrificing
any expected return. This was the cental insight of Markowitz who (in his framework) recognized that investors
seek to minimize variance for a given level of expected return or, equivalently, they seek to maximize expected
return for a given constraint on variance.
Before formulating and solving the mean variance problem consider Figure 1 below. There were n = 8 securities
with given mean returns, variances and covariances. We generated m = 200 random portfolios from these n
securities and computed the expected return and volatility, i.e. standard deviation, for each of them. They are
plotted in the figure and are labelled “inefficient”. This is because every one of these random portfolios can be
improved. In particular, for the same expected return it is possible to find an “efficient” portfolio with a smaller
volatility. Alternatively, for the same volatility it is possible to find an efficient portfolio with higher expected
return.

1.1 The Efficient Frontier without a Risk-free Asset

We will consider first the mean-variance problem when a risk-free security is not available. We assume that
there are n risky securities with the corresponding return vector, R, satisfying
R ∼ MVNn (µ, Σ). (1)
Mean-Variance Optimization and the CAPM 2

Figure 1: Sample Portfolios and the Efficient Frontier (without a Riskfree Security).

The mean-variance portfolio optimization problem is formulated as:

1 0
min w Σw (2)
w 2

subject to w0 µ = p
and w0 1 = 1.

Note that the specific value of p will depend on the risk aversion of the investor. This is a simple quadratic
optimization problem and it can be solved via standard Lagrange multiplier methods.

Exercise 1 Solve the mean-variance optimization problem (2).

When we plot the mean portfolio return, p, against the corresponding minimal portfolio volatility / standard
deviation we obtain the so-called portfolio frontier. We can also identify the portfolio having minimal variance
among all risky portfolios: this is called the minimum variance portfolio. The points on the portfolio frontier
with expected returns greater than the minimum variance portfolio’s expected return, R̄mv say, are said to lie on
the efficient frontier. The efficient frontier is plotted as the upper blue curve in Figure 1 ar alternatively, the blue
curve in Figure 2.

Exercise 2 Let w1 and w2 be (mean-variance) efficient portfolios corresponding to expected returns r1 and
r2 , respectively, with r1 6= r2 . Show that all efficient portfolios can be obtained as linear combinations of w1
and w2 .

The result of the previous exercise is sometimes referred to as a 2-fund theorem.

1.2 The Efficient Frontier with a Risk-free Asset

We now assume that there is a risk-free security available with risk-free rate equal to rfP . Let
n
w := (w1 , . . . , wn )0 be the vector of portfolio weights on the n risky assets so that 1 − i=1 wi is the weight on
the risk-free security. An investor’s portfolio optimization problem may then be formulated as
1 0
min w Σw (3)
w 2
n
!
X
subject to 1− wi rf + w0 µ = p.
i=1
Mean-Variance Optimization and the CAPM 3

Figure 2: The Efficient Frontier with a Riskfree Security.

The optimal solution to (3) is given by

w = ξ Σ−1 (µ − rf 1) (4)
2 2
where ξ := σmin /(p − rf ) and σmin is the minimized variance, i.e., twice the value of the optimal objective
function in (3). It satisfies
2 (p − rf )2
σmin = (5)
(µ − rf 1)0 Σ−1 (µ − rf 1)
where 1 is an n × 1 vector of ones. While ξ (or p) depends on the investor’s level of risk aversion it is often
inferred from the market portfolio. For example, if we take p − rf to denote the average excess market return
2 2
and σmin to denote the variance of the market return, then we can take σmin /(p − rf ) as the average or market
value of ξ.
Suppose now that rf < R̄mv . When we allow our portfolio to include the risk-free security the efficient frontier
becomes a straight line that is tangential to the risky efficient frontier and with a y-intercept equal to the
risk-free rate. This is plotted as the red line in Figure 2. That the efficient frontier is a straight line when we
include the risk-free asset is also clear from (5) where we see that σmin is linear in p. Note that this result is a
1-fund theorem since every investor will optimally choose to invest in a combination of the risk-free security
and a single risky portfolio, i.e. the tangency portfolio. The tangency portfolio, w∗ , is given by the optimal
w of (4) except that it must be scaled so that its component sum to 1. (This scaled portfolio will not depend
on p.)

Exercise 3 Without using (5) show that the efficient frontier is indeed a straight line as described above.
Hint: consider forming a portfolio of the risk-free security with any risky security or risky portfolio. Show that
the mean and standard deviation of the portfolio varies linearly with α where α is the weight on the
risk-free-security. The conclusion should now be clear.

Exercise 4 Describe the efficient frontier if no borrowing is allowed.

The Sharpe ratio of a portfolio (or security) is the ratio of the expected excess return of the portfolio to the
portfolio’s volatility. The Sharpe optimal portfolio is the portfolio with maximum Sharpe ratio. It is
straightforward to see in our mean-variance framework (with a risk-free security) that the tangency portfolio,
w∗ , is the Sharpe optimal portfolio.

1.3 Including Portfolio Constraints

We can easily include linear portfolio constraints in the problem formulation and still easily solve the resulting
quadratic program. No-borrowing or no short-sales constraints are examples of linear constraints as are leverage
Mean-Variance Optimization and the CAPM 4

and sector constraints. While analytic solutions are generally no longer available, the resulting problems are still
easy to solve numerically. In particular, we can still determine the efficient frontier.

1.4 Weaknesses of Traditional Mean-Variance Analysis

The traditional mean-variance analysis of Markowitz has many weaknesses when applied naively in practice.
They include:
1. The tendency to produce extreme portfolios combining extreme shorts with extreme longs. As a result,
portfolio managers generally do not trust these extreme weights. This problem is typically caused by
estimation errors in the mean return vector and covariance matrix.

Figure 3: The Efficient Frontier, Estimated Frontiers and Realized Frontiers.

Consider Figure 3, for example, where we have plotted the same efficient frontier (of risky securities) as in
Figure 2. In practice, investors can never compute this frontier since they do not know the true mean
vector and covariance matrix of returns. The best we can hope to do is to approximate it. But how might
we do this? One approach would be to simply estimate the mean vector and covariance matrix using
historical data. Each of the black dashed curves in Figure 3 is an estimated frontier that we computed by:
(i) simulating m = 24 sample returns from the true (in this case, multivariate normal) distribution (ii)
estimating the mean vector and covariance matrix from this simulated data and (iii) using these estimates
to generate the (estimated) frontier. Note that the blue curve in Figure 3 is the true frontier computed
using the true mean vector and covariance matrix.

The first observation is that the estimated frontiers are quite random and can differ greatly from the true
frontier. They may lie below or above the true frontier or they may cross it and an investor who uses such
an estimated frontier to make investment decisions may end up choosing a poor portfolio. How poor? The
dashed red curves in Figure 3 are the realized frontiers that depict the true portfolio mean - volatility
tradeoff that results from making decisions based on the estimated frontiers. In contrast to the estimated
frontiers, the realized frontiers must always (why?) lie below the true frontier. In Figure 3 some of the
realized frontiers lie very close to the true frontier and so in these cases an investor would do very well.
But in other cases the realized frontier is far from the (generally unobtainable) true efficient frontier.

These examples serve to highlight the importance of estimation errors in any asset allocation procedure.
Note also that if we had assumed a heavy-tailed distribution for the true distribution of portfolio returns
then we might expect to see an even greater variety of sample mean-standard deviation frontiers. In
addition, it is worth emphasizing that in practice we may not have as many as 24 relevant observations
available. For example, if our data observations are weekly returns, then using 24 of them to estimate the
joint distribution of returns is hardly a good idea since we are generally more concerned with estimating
Mean-Variance Optimization and the CAPM 5

conditional return distributions and so more weight should be given to more recent returns. A more
sophisticated estimation approach should therefore be used in practice. More generally, it must be stated
that estimating expected returns using historical data is very problematic and is not advisable!

2. The portfolio weights tend to be extremely sensitive to very small changes in the expected returns. For
example, even a small increase in the expected return of just one asset can dramatically alter the optimal
composition of the entire portfolio. Indeed let w and wb denote the true optimal and estimated optimal
portfolios, respectively, corresponding to the true mean return vector, µ, and the sample mean return
vector, µ
b, respectively. Then Best and Grauer (1991) showed that
 
1 γmax
||w − w||
b ≤ ξ ||µ − µ b|| 1+
γmin γmin

where γmax and γmin are the largest and smallest eigen values, respectively, of the covariance matrix, Σ.
Therefore the sensitivity of the portfolio weights to errors in the mean return vector grows as the ratio
γmax /γmin grows. But this ratio, when applied to the estimated covariance matrix, Σ, b typically becomes
large as the number of asset increases and the number of sample observations is held fixed. As a result,
we can expect large errors for large portfolios with relatively few observations.
3. While it is commonly believed that errors in the estimated means are of much greater significance, errors
in estimated covariance matrices can also have considerable impact. While it is generally easier to
estimate covariances than means, the presence of heavy tails in the return distributions can result in
significant errors in covariance estimates as well. These problems can be mitigated to varying extents
through the use of more robust estimation techniques.
As a result of these weaknesses, portfolio managers traditionally have had little confidence in mean-variance
analysis and therefore applied it very rarely in practice. Efforts to overcome these problems include the use of
better estimation techniques such as the use of shrinkage estimators, robust estimators and Bayesian techniques
such as the Black-Litterman framework introduced in the early 1990’s. (In addition to mitigating the problem of
extreme portfolios, the Black-Litterman framework allows users to specify their own subjective views on the
market in a consistent and tractable manner.) Many of these techniques are now used routinely in general asset
allocation settings. It is worth mentioning that the problem of extreme portfolios can also be mitigated in part
by placing no short-sales and / or no-borrowing constraints on the portfolio.

Figure 4: Robust Estimation of the Efficient Frontier.

Mean-Variance Optimization and the CAPM 6

In Figure 4 above we have shown an estimated frontier that was computed using a more robust estimation
procedure. We see that it lies much closer to the true frontier which is also the case with it’s corresponding
realized frontier.

2 The Capital Asset Pricing Model (CAPM)

If every investor is a mean-variance optimizer then we can see from Figure 2 and our earlier discussion that each
of them will hold the same tangency portfolio of risky securities in conjunction with a position in the risk-free
asset. Because the tangency portfolio is held by all investors and because markets must clear, we can identify
this portfolio as the market portfolio. The efficient frontier is then termed the capital market line.
Now let Rm and R̄m denote the return and expected return, respectively, of the market, i.e. tangency, portfolio.
The central insight of the Capital Asset-Pricing Model is that in equilibrium the riskiness of an asset is not
measured by the standard deviation of its return but by its beta. In particular, there is a linear relationship
between the expected return, R̄ = E[R] say, of any security (or portfolio) and the expected return of the market
portfolio. It is given by
R̄ = rf + β (R̄m − rf ) (6)
where β := Cov(R, Rm )/Var(Rm ). In order to prove (6), consider a portfolio with weights α and weight 1 − α
on the risky security and market portfolio, respectively. Let Rα denote the (random) return of this portfolio as a
function of α. We then have

E [Rα ] = αR̄ + (1 − α)R̄m (7)

2
σR α
= α 2 σR
2
+ (1 − α)2 σR
2
m
+ 2α(1 − α)σR,Rm (8)
2 2 2
where σR α
, σR and σR m
are the returns variances of the portfolio, security and market portfolio, respectively.
We use σR,Rm to denote Cov(R, Rm ). Now note that as α varies, the mean and standard deviation,
2
(E [Rα ] , σR α
), trace out a curve that cannot (why?) cross the efficient frontier. This curve is depicted as the
dashed curve in Figure 5 below.

Figure 5: Proving the CAPM relationship.

Therefore at α = 0 this curve must be tangent to the capital market line. Therefore the slope of the curve at
Mean-Variance Optimization and the CAPM 7

α = 0 must equal the slope of the capital market line. Using (7) and (8) we see the former slope is given by
  
d E[Rα ]  d E[Rα ] d σRα 
=
d σRα α=0 dα d α α=0


σRα R̄ − R̄m 
=

2 − (1 − α)σ 2
ασR Rm + (1 − 2α)σR,Rm α=0



σRm R̄ − R̄m
= 2 . (9)
−σR m
+ σR,Rm


The slope of the capital market line is R̄m − rf /σRm and equating the two therefore yields


σRm R̄ − R̄m R̄m − rf
2 = (10)
−σR m
+ σR,R m
σRm

which upon simplification gives (6).

The CAPM result is one of the most famous results in all of finance and, even though it arises from a simple
one-period model, it provides considerable insight to the problem of asset-pricing. For example, it is well-known
that riskier securities should have higher expected returns in order to compensate investors for holding them.
But how do we measure risk? Counter to the prevailing wisdom at the time the CAPM was developed, the
riskiness of a security is not measured by its return volatility. Instead it is measured by its beta which is
proportional to its covariance with the market portfolio. This is a very important insight. Nor, it should be
noted, does this contradict the mean-variance formulation of Markowitz where investors do care about return
variance. Indeed, we derived the CAPM from mean-variance analysis!

Exercise 5 Why does the CAPM result not contradict the mean-variance problem formulation where investors
do measure a portfolio’s risk by its variance?

The CAPM is an example of a so-called 1-factor model with the market return playing the role of the single
factor. Other factor models can have more than one factor. For example, the Fama-French model has three
factors, one of which is the market return. Many empirical research papers have been written to test the CAPM.
Such papers usually perform regressions of the form

Ri − rf = αi + βi (Rm − rf ) + i (11)

where αi (not to be confused with the α we used in the proof of (6)) is the intercept and i is the idiosyncratic
or residual risk which is assumed to be independent of Rm and the idiosyncratic risk of other securities. If the
CAPM holds then we should be able to reject the hypothesis that αi 6= 0. The evidence in favor of the CAPM is
mixed. But the language inspired by the CAPM is now found throughout finance. For example, we use β’s to
denote factor loadings and α’s to denote excess returns even in non-CAPM settings.