Mean-Variance Analysis and the CAPM

Dr. Markus Reisinger Ludwig-Maximilians-Universitat M

unchen
Theorie der Finanzmarkte Sommersemester 2008

The Mean-Variance Problem

Suppose an individual can invest in N risky assets, labelled n = 1, . . . , N , with the following
n with expected value
characteristics. Asset n has the random gross return R
n ] = Rn
E[R
and variance
n ] = n2 > 0.
Var[R
The covariance between assets n and m is
n, R
m ] = nm ;
Cov[R
note that nn = n2 . These characteristics are summarised in the vector of expected returns
R1
..
R= .
RN

and the variance-covariance matrix

11 1N
..
.. .
V= .
.
N 1 N N

The investors portfolio can be described by a vector

w1
..
w= .
wN

where wn specifies the proportion of wealth invested in asset n. We assume that all of the
individuals wealth is invested, so the portfolio weights have to satisfy the constraint
N
X

wn = 1.

n=1

c Sven Rady 2008

Mean-Variance Analysis & CAPM

Dr. M. Reisinger

Apart from this budget identity, we do not impose any further constraints on the portfolio
choice. This supposes for example that assets are perfectly divisible in the sense that the
investor can take an arbitrarily small position in any asset (such as one cents worth of IBM
stock). More striking, a portfolio weight wn can even be negative. This corresponds to a
so-called short position in a security and is achieved by selling the asset short: first, you
borrow the security (from your friendly stock broker, say) and sell it to somebody else; in the
next period, you repurchase the security on the market, and hand it back to your broker. If
the price of the security falls in the meantime, you end up with a nice profit!1
The expected return and the variance of a portfolio w are given by
Rw =

N
X

wn Rn

n=1

and
2
w

N
X

wn wm nm .

n,m=1

Mean-variance analysis goes back to Markowitz (1952). Its principal assumption is that the
investor bases the portfolio decision solely on these two characteristics of the uncertain portfolio return. More specifically, it is postulated that the agents preferences induce the favouring
of higher means and lower variances. The class of potentially optimal portfolios for the investor are therefore those with the greatest expected return for a given level of variance and,
simultaneously, the smallest variance for a given expected return. Such portfolios are termed
mean-variance efficient. The set of all these portfolios is called the efficient frontier.
For reasons of tractability, we shall first study the larger set of minimum-variance portfolios,
i.e., portfolios that achieve their expected return with the smallest possible variance. All meanvariance efficient portfolios are also minimum-variance portfolios, but the converse is not true.
Thus the efficient frontier is a strict subset of the minimum-variance set. Anticipating the main
results of this section, Figure ?? shows the efficient frontier and the minimum-variance set as
they will appear in the variance/mean plane with coordinates ( 2 , R). The arrows indicate the
preferred directions, that is, the directions in which the investor would like to move.
To calculate the minimum-variance portfolio for a given expected return R, we have to solve
the following optimisation problem:2
min

wIRN

1
2

N
X
n,m=1

wn wm nm

s.t.

N
X
n=1

wn = 1 and

N
X

wn Rn = R.

n=1

We shall first look at a numerical example with three assets, and then consider the general
case.

In reality, only some investors are allowed to sell securities short, and they have to make considerable margin
payments as collateral for the borrowed securities. Moreover, the short-seller forgoes any dividend payments on
the borrowed shares. We neglect these complications here, and allow unlimited short sales.
2
Multiplying the portfolio variance by 12 does not change the solution, but makes first order conditions look
nicer.

Mean-Variance Analysis & CAPM

Dr. M. Reisinger

R
6

.....
...................
...................
...................
...................
...................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.....
...................
...................
.............
.............
.............
.
.
.
.
.
.
.
.
.
.
.
.
...
.............
.............
.............
.............
........
.
.
.
.
.
.
.
.
.
.........
........
.........
.........
........
.
.
.
.
.
.
.
.
.....
.....
.....
....
....
.
.
.
...
..
....
...
...
...
....
....
.....
.....
.....
.......
........
.........
.........
.........
.........
.........
.........
............
.............
.............
.............
.............
.............
.............
.............
...............
...................
...................
...................
...................
...................
...................
...................
...................
...

- 2

Figure 1: Minimum-variance set and efficient frontier

1.1

An Example with N = 3

Suppose we have three assets n = 1, 2, 3 with expected returns

R1
1

R = R2 = 2
3
R3
and the following variances and covariances:

12 12 13
1 1 0

2
V = 21 2 23 = 1 4 0 .
31 32 32
0 0 9
Thus, the return on asset 3 is uncorrelated with the returns on assets 1 and 2, and these two
1
returns have the correlation coefficient 12 = 112
2 = 2 .
The investor has to choose a portfolio

w1

w = w2
w3
where wn stands for the fraction of wealth invested in asset n, and w1 + w2 + w3 = 1.
2 + w3 R
3 with expectation
w = w1 R
1 + w2 R
The risky return on such a portfolio is R
w ] = w1 R1 + w2 R2 + w3 R3 = w1 + 2w2 + 3w3
Rw = E[R

Mean-Variance Analysis & CAPM

Dr. M. Reisinger

and variance
2
w ]
w
= Var[R
w , R
w ]
= Cov[R

3]
2 + w3 R
1 + w2 R
3 , w1 R
2 + w3 R
1 + w2 R
= Cov[w1 R
= w12 12 + w22 22 + w32 32 + 2w1 w2 12 + 2w1 w3 13 + 2w2 w3 23
= w12 + 4w22 + 9w32 + 2w1 w2 .
Now suppose the investor wants to achieve an expected portfolio return of Rw = R with as
2 as possible. This amounts to solving the following problem:
little portfolio risk w
min 1 [w12
w1 ,w2 ,w3 2

+ 4w22 + 9w32 + 2w1 w2 ]

s.t. w1 + w2 + w3 = 1 and w1 + 2w2 + 3w3 = R.

The strict convexity of the objective function and the linearity of the constraints imply that
there exists a unique solution. The corresponding Lagrangian is
L = 12 [w12 + 4w22 + 9w32 + 2w1 w2 ] + [1 w1 w2 w3 ] + [R w1 2w2 3w3 ].
Differentiating with respect to the portfolio weights w1 , w2 , w3 yields the first order conditions
w1 + w2 = 0,
4w2 + w1 2 = 0,
9w3 3 = 0.
The third condition immediately tells us that
w3 =

+ .
9
3

Subtracting the first condition from the second implies 3w2 = 0 and so

w2 = .
3
Inserting this into the first condition, we then get
w1 = + w2 = +

2
.
3

In the usual way, we now solve for the Lagrange multipliers using the two constraints on the
portfolio weights:
2
10 4
+ + +
=
+
,
3
3
9
3
9
3
2 2
4 7
= +
+
+ + =
+
.
3
3
3
3
3

1 = w1 + w2 + w3 = +
R = w1 + 2w2 + 3w3

This is a system of two linear equations for two unknowns:

10 + 12 = 9,
4 + 7 = 3R.
There are many different ways to solve this system, all leading to the solution
=
=

63 36R
,
22
15R 18
30R 36
=
.
11
22

Mean-Variance Analysis & CAPM

Dr. M. Reisinger

Inserting these values in the above expressions for the portfolio weights, we finally obtain the
minimum-variance portfolio that achieves the desired expected return R with the smallest
possible risk:
39 16R
2
=
,
3
22
10R 12
=
,
22

6R 5
+
=
.
3
22

w1 = +
w2 =
w3 =

Let us calculate a few concrete numbers:

R

w1

w2

w3

23
22

2
22

1
22

7
22

8
22

7
22

9
22

18
22

13
22

Note that the minimal-variance portfolios for R = 1 and 3 involve short selling. Moreover, the
best way to earn an expected return of R = 1, say, is not simply to hold asset 1, but to combine
it with the other assets. This way, the investor reaps the benefits of diversification.
To see this more clearly, let us calculate the minimal variance 2 with which an expected return
of R can be achieved. This is precisely the variance of the optimal portfolio that we computed
above, so we get
2 = w12 + 4w22 + 9w32 + 2w1 w2
i
1 h
2
2
2
=
(39

16R)
+
4(10R

12)
+
9(6R

5)
+
2(39

16R)(10R

12)
222
660R2 1584R + 1386
=
222
2
30R 72R + 63
.
=
22
For the above three values of R, we find the following minimal variances and standard deviations:
R
1
2
3

2
0.95
1.77
5.32

0.98
1.33
2.30

In each case, the optimal minimal-variance portfolio achieves the given expected return at lower
risk than the primitive asset with the same expected return. This is the result of diversification:
as the three assets are imperfectly correlated, combining them reduces the overall risk.

Mean-Variance Analysis & CAPM

Dr. M. Reisinger

How low can we push the portfolio variance? The answer is the global minimum variance g2 ,
which we find by solving the first order condition
d 2
= 60R 72 = 0.
dR
Thus, portfolio variance is globally minimal at Rg = 65 , and the global minimum variance equals
9
g2 = 10
. The corresponding global minimum-variance portfolio g is given by the portfolio
9
1
weights g1 = 10
, g2 = 0 and g3 = 10
.
In the variance/mean plane with coordinates ( 2 , R), the equation
2 =

30R2 72R + 63
22

that we derived for minimum-variance portfolios describes a parabola lying on its side as depicted in Figure ??. In the standard deviation/mean plane with coordinates (, R), the locus
of the minimum-variance portfolios is a hyperbola as shown in Figure ??.

R
............
.................
................
.................
................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
....
.................
.................
................
.................
.................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.....
..............
..............
...............
...............
..........
.
.
.
.
.
.
.
.
.
.
.
..
.............
.............
...........
............
..........
.
.
.
.
.
.
.
.
...
.........
........
......
.......
......
.
.
.
.
...
....
...
.... ...... ...... ...... ...... .....
...
... .......
.....
.
.. .............
.
.......
......
........
.
.........
..
...........
............
.
............
..
.............
.............
............
...
...........
...............
...............
...
...............
..............
...............
..
.................
.................
.................
..
.................
.................
.................
.................
..
.................
.
.................
.................
.
.................
..
....
..
..

Rg

tg

Figure 2: The minimum-variance hyperbola

The efficient frontier corresponds to the part of the hyperbola which lies above the apex of the
curve, and the apex corresponds to the global minimum-variance portfolio g. Using the above
values of Rg and g2 , we can rewrite the equation for the minimum variance more simply as
2
w

30
6
=
R
22
5


2

9
.
10

Mean-Variance Analysis & CAPM

1.2

Dr. M. Reisinger

The General Solution to the Mean-Variance Problem

In order to find out how the example generalises to an arbitrary number of assets with arbitrary
expected returns, variances and covariances, we have to solve the mean-variance problem in its
general form. This is most conveniently done in vector notation.3
Throughout, we shall assume that the expected returns vector R is not a multiple of the vector
1 whose components are all equal to 1. This rules out the uninteresting case where all assets
have the same expected return.
We shall also assume that the symmetric matrix V is positive definite, i.e., w> Vw > 0 for
all w IRN with w 6= 0. This implies that V is invertible, and guarantees that no asset is
redundant in the sense that its return can be expressed as a linear combination of the returns
of other assets.
In vector notation, the restriction that the weights of a portfolio w add up to 1 can be written
as
1> w = 1.
The expected return and the variance of a portfolio w can be written as
Rw = w > R
and
2
w
= w> Vw.

Using these expressions, we can write the investors problem as

min

wIRN

1 >
2 w Vw

s.t. 1> w = 1 and R> w = R.

Differentiating the Lagrangian

L = 21 w> Vw + (1 1> w) + (R R> w)
with respect to the portfolio weights, we get the following first order condition in vector form:
Vw 1 R = 0.
Thus the solution is
w = V1 1 + V1 R.
To solve for and , we plug this expression into the two constraints:
1 = 1> w = 1> V1 1 + 1> V1 R
and
R = R> w = R> V1 1 + R> V1 R.
After introducing the constants
A = 1> V1 1,
B = 1> V1 R,
C = R> V1 R,
3

As an exercise, try to redo the above example in vector notation. This will enable you to compare the two
methods. Of course, either will be fine in the exam.

Mean-Variance Analysis & CAPM

Dr. M. Reisinger

we are left with the following 2 2 system of linear equations:

A + B = 1,
B + C = R.
The solution is
=
=

C BR
,

AR B

with
= AC B 2 ,
provided 6= 0. To show that this is indeed the case, recall that the inverse of a positive
definite matrix is itself positive definite. In particular, we have A > 0 and C > 0. Moreover,
note that
(BR C1)> V1 (BR C1) = B 2 C 2B 2 C + AC 2 = C.
The left hand side is positive since BR C1 6= 0. By the fact that C > 0, we therefore have
> 0.
We can now derive an equation for the smallest variance 2 compatible with an expected return
of R. This is most easily done by premultiplying the first order condition with w> :
0 = w> Vw w> 1 w> R = 2 R.
Rearranging and inserting the above expressions for and , we finally obtain
2 =

AR2 2BR + C
.

Exactly as in the example, ( 2 , R), this equation describes a parabola in the variance/mean
plane with coordinates ( 2 , R), and a hyperbola in the standard deviation/mean plane with
coordinates (, R).
The efficient frontier corresponds to the part of the hyperbola which lies above the apex of the
curve. The apex can be identified with the global minimum-variance portfolio g. This portfolio
is obtained by setting , the Lagrange multiplier for the constraint R> w = R, equal to zero
(why?), which implies = 1/A, an expected return of B/A, and a variance of 1/A. Inserting
these values for the Lagrange multipliers into the formula for optimal portfolios, we obtain the
representation
1
V1 1
g = V1 1 = > 1
A
1 V 1
for the global minimum-variance portfolio.
We can summarise these findings in
Proposition 1 A portfolio is in the minimum-variance set if and only if its expected return R
and variance 2 satisfy
A
2 = (R Rg )2 + g2

2
where Rg = B/A and g = 1/A are the expected return and variance of the global minimumvariance portfolio. The efficient frontier consists of those minimum-variance portfolios that
have an expected return of at least Rg .

Mean-Variance Analysis & CAPM

Dr. M. Reisinger

Two-Fund Separation

The previous section has shown that the minimum-variance portfolio for an expected return of
R is
w(R) = (R)V1 1 + (R)V1 R
with the linear functions (R) and (R) as calculated above. In particular, this means that
the minimum-variance portfolios form a straight line in IRN which is parameterised by the
expected return R.
An interesting financial implication of this fact is a phenomenon called two-fund separation.
Fix two distinct values R1 6= R2 . For any R, we can then find a unique real number such
that
R = R1 + (1 )R2 .
You can check very easily that this implies
(R) = (R1 ) + (1 )(R2 ),
(R) = (R1 ) + (1 )(R2 ),
and therefore
w(R) = w(R1 ) + (1 )w(R2 ).
Thus, all investors who choose portfolios by examining only mean and variance can be satisfied
by holding different combinations of just two mutual funds. According to the above analysis,
the choice of the two funds does not matter as long as both are in the minimum-variance set.
Proposition 2 The minimum-variance set (and hence the efficient frontier) can be generated
by forming combinations of any two distinct minimum-variance portfolios.

Introducing a Risk-Free Security

Consider an additional security, labelled n = 0, which yields a safe gross return of R0 > 1.
We assume that this asset can be bought and sold in unlimited amounts, that is, unlimited
borrowing and lending are possible at the interest rate r = R0 1.
As before, let R IRN and V IRN N denote the vector of expected returns and the variance
covariance matrix of the risky assets. The proportions of wealth invested in the risky assets will
again be described by a vector w IRN . The difference to the previous case without risk-free
asset is that the holdings of risky assets are now unconstrained since we can always choose the
riskless investment so as to satisfy the budget constraint:
w0 = 1

N
X

wn = 1 1> w

n=1

where 1 is again the vector whose N components are all equal to 1. In other words, a portfolio
in the N + 1 assets n = 0, 1, . . . , N is fully specified once we know the holdings of the N risky
assets n = 1, . . . , N , i.e., the vector w. The expected overall portfolio return corresponding to
the risky investment w is
Rw =

N
X
n=1

wn R0 +

N
X
n=1

wn Rn = R0 +

N
X
n=1

wn (Rn R0 )

Mean-Variance Analysis & CAPM

10

Dr. M. Reisinger

or, in vector notation,

Rw = (1 1> w)R0 + R> w = R0 + (R R0 1)> w.
As the covariances with the risk-free asset are all zero, the portfolio variance is still
2
w

N
X

wn wm nm = w> Vw.

n,m=1

The minimum-variance portfolio for a given expected return of R is the solution to the following
optimisation problem:
min

wIRN

1 >
2 w Vw

s.t. (R R0 1)> w = R R0 .

Differentiating the Lagrangian

L = 21 w> Vw + (R R0 (R R0 1)> w)
with respect to the portfolio weights, we get the following first order condition for an optimum:
Vw (R R0 1) = 0.
Thus the solution is
w = V1 (R R0 1).
To solve for , we plug this expression into the constraint:
R R0 = (R R0 1)> w = (R R0 1)> V1 (R R0 1),
hence
=

R R0
H

where
H = (R R0 1)> V1 (R R0 1) = C 2BR0 + AR02 > 0.
(Why is H > 0?)
The smallest variance compatible with the expected return R can be obtained by premultiplying
the first order condition with w> :
0 = w> Vw w> (R R0 1) = 2 (R R0 )
and so

(R R0 )2
.
H
In the variance/mean plane, this equation describes again a parabola. In the standard deviation/mean plane, however, the situation is now considerably simpler. Taking square roots on
both sides of the last equation, we obtain
2 =

|R R0 |

.
H

In (, R)-space, this describes a pair of rays with common intercept at R0 and slopes of H.
Proposition 3 In the presence of a risk-free asset with return R0 , a portfolio belongs to the
minimum-variance set if and only if its expected return R and standard deviation satisfy
|R R0 |

.
H
The efficient frontier consists of those minimum-variance portfolios that have an expected return
of at least R0 .
=

Mean-Variance Analysis & CAPM

11

Dr. M. Reisinger

Covariances and Expected Returns

The covariances of arbitrary portfolios with portfolios in the minimum-variance set will turn
out to be of great importance.
Consider a portfolio w in the minimum-variance set and an arbitrary portfolio p. Premultiplying the first order condition
Vw (Rw )(R R0 1) = 0
with w> and p> , respectively, we get
2
w
= (Rw )(Rw R0 )

and
pw = (Rw )(Rp R0 ).
Eliminating (Rw ), we find the following expression for the expected return on p:
Rp = R 0 +

pw
(Rw R0 ).
2
w

Thus, the expected return on any portfolio p depends linearly on its covariance with a given
minimum-variance portfolio w. It is customary to write this linear relationship as
Rp = R0 + pw (Rw R0 )
where
pw =

pw
2
w

is the beta of portfolio p with respect to portfolio w. Note that pw is the weighted average of
the asset betas nw :
N
X

pw =

pn nw .

n=1

You are asked in an exercise to show that nw measures the contribution of security n to the
risk of portfolio w in the following sense:
nw =

1 w
.
w wn

To gain some intuition for the above linear relationship between the expected return of a
portfolio and its beta relative to a minimum-variance portfolio, we start with the special case
p is uncorrelated with R
w and so pw = 0. For any real number x, consider the portfolio
where R
2 +x2 2 .
w(x) = w+xp. Its expected return is R(x) = Rw +x(Rp R0 ), its variance 2 (x) = w
p
2
Note that at x = 0, the derivative of (x) vanishes, so adding or subtracting the first marginal
unit of portfolio p does not change the overall portfolio risk. If Rp R0 6= 0, however, there is
a first-order effect on the expected portfolio return. This suggests that unless the excess return
on p is zero, there are ways to improve upon portfolio w by perturbing it in the direction of
portfolio p.
To see this more precisely, we suppose Rw 6= R0 and redefine the perturbed portfolio so as to
keep its expected return constant:


w(x) = 1 x

R p R0
w + xp.
Rw R0


Mean-Variance Analysis & CAPM

12

Dr. M. Reisinger

Then,
R p R0
(Rw R0 ) + x(Rp R0 ) = Rw
R(x) = R0 + 1 x
Rw R0


and
R p R0
(x) = 1 x
Rw R0
2

2

2
w
+ x2 p2 .

2 (R R )/(R R ), so if R 6= R , we
The derivative of this variance at x = 0 equals 2w
p
0
w
0
p
0
can find portfolios w(x) that earn the same expected return as w but have lower variance. For
w to be a minimum-variance portfolio, therefore, any portfolio whose return is uncorrelated
with that of w must earn the riskfree rate of return. This is exactly what the above linear
relationship predicts in the case that beta is zero.

The intuition for this special case can be summarized as follows. At the margin, a portfolio
that is uncorrelated with a given portfolio does not add risk to the latter, hence is a perfect
substitute for the riskless asset. If its expected return differs from the riskless one, the investor
can exploit this to improve upon the given portfolio.
p and R
w are correlated can be reduced to this special case. In fact, we
The general case where R
p = + R
w +  with real constants and
can apply a least-squares decomposition and write R
2
= pw /w (which is precisely pw as defined above) and a random variable  uncorrelated
w . Now, the return on the portfolio q = p w equals R
q = R
p (R
w R0 ) =
with R
w . If w is a minimum-variance portfolio, therefore,
+ R0 +  and so is uncorrelated with R
the expected return on q must be R0 , which is the same as Rp = R0 + (Rw R0 ).

The Tangency Portfolio

Two-fund separation holds by virtue of exactly the same arguments as in the case without
a risk-free asset. Again, the minimum-variance set can be generated from any two distinct
minimum-variance portfolios. In the present situation, however, there is a natural choice of
funds: the riskless asset and a particular portfolio with risky investments only. Clearly, the risky
investment part of any minimum-variance portfolio is a multiple of V1 (R R0 1). Calculating
1> V1 (R R0 1) = B AR0 ,
we see that there are two cases, depending on whether R0 is different from Rg = B/A or not.
If R0 6= Rg , then
t=

V1 (R R0 1)
B AR0

defines the unique minimum-variance portfolio with 1> t = 1, hence no investment in the
risk-free asset. This is the minimum-variance portfolio for = 1/(B AR0 ), which implies
Rt = R 0 +

H
C BR0
=
B AR0
B AR0

and
t2 =

H
.
(B AR0 )2

Since this portfolio involves the risky assets only, the point corresponding to t in the (, R)plane cannot lie to the left of the risky-asset minimum-variance hyperbola. Nor can it lie to

Mean-Variance Analysis & CAPM

13

Dr. M. Reisinger

the right of the hyperbola because this would mean that t does not achieve the lowest possible
variance among all portfolios with expected return Rt . So (t , Rt ) must lie on the risky-asset
minimum-variance hyperbola. In other words, t belongs to the risky-asset minimum-variance
set.4
At any other point than (t , Rt ), the risky-asset minimum-variance hyperbola must lie to the
right of the overall minimum-variance locus. (Why?) So the point corresponding to t in the
(, R)-plane is in fact the unique tangency point between the hyperbola and a ray starting from
the intercept R0 . The portfolio t is therefore called the tangency portfolio.5
We have shown:
Proposition 4 In the presence of a risk-free asset with return R0 6= Rg , the minimum-variance
set can be generated by combining the risk-free asset with the tangency portfolio. This portfolio
is the only minimum-variance portfolio that involves no investment in the risk-free asset, and
is determined by the point of tangency between the risky-asset minimum-variance hyperbola and
a ray with intercept R0 in (, R)-space.
Let us now turn to the efficient frontier.
Proposition 5 If R0 < Rg , then Rt > Rg , so the tangency occurs on the upper limb of the
risky-asset minimum-variance hyperbola. In this case, the efficient frontier is generated by
combining a long (or zero) position in the tangency portfolio with riskless lending or borrowing.
If R0 > Rg , the tangency occurs on the lower limb of the hyperbola (Rt < Rg ), so the tangency
portfolio is inefficient. The efficient frontier is generated by combining a short (or zero) position
in the tangency portfolio with riskless lending.
One can prove this result by calculating the product (Rt Rg )(Rg R0 ) and verifying that it
is always positive for R0 6= Rg . The case R0 < Rg is illustrated in Figure ??.
The intuition behind the tangency result for this case should become clear from the diagram.
In fact, by combining any risky-asset minimum-variance portfolio w with risk-free borrowing or
lending, the investor can generate a ray from the intercept R0 through the point corresponding
to w (see the dashed line). Indeed, investing the proportion > 0 of ones wealth in w and
the proportion 1 in the risk-free asset yields an expected return of R = R0 + (Rw R0 )
with a standard deviation of = w . Thus, parameterises a ray with equation
R = R0 +

Rw R0
.
w

Its slope, (Rw R0 )/w , is called the Sharpe ratio of portfolio w. To construct the efficient
frontier, the investor aims at the highest possible expected return for a given standard deviation.
This means pushing up the slope of the ray as far as possible, that is, until the ray becomes
tangent. The investor can then chose the most preferred portfolio p on the tangent line.
4

Formally, one can see this by writing

t=

R0
1
V1 1 +
V1 R
B AR0
B AR0

which is the same as t = V1 1 + V1 R with the Lagrange multipliers = R0 /(B AR0 ) = (Rt ) and
= 1/(B AR0 ) = (Rt ).
5
From the previous footnote, we see that the expected return on the tangency portfolio satisfies (Rt ) =
(Rt ). This identity fully characterizes the tangency portfolio. You should be able to see why after recalling that
a Lagrange multiplier measures the effect on the objective function (here: the portfolio variance) of a marginal
change in the constraint (here: the target return).

Mean-Variance Analysis & CAPM

14

Dr. M. Reisinger

...
....
.....
....
........
.........
....
...
.........
.
.
.
.
.
.
.
.
.
..
....
.........
......
........
......
........
...
......
.........
.................
......
.........
.................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... ..
...
.........
.................
.....
........
.................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
..
...... .......................
.........
........
..............................
........
............................
.........
.................
.........
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
....
.................
.........
........ .................................. ..
........
. .
......... ............. ....
............................... ..... .....
.
.
.
.
.
.
.
.. ...
...................... ..... .....
...................
............. .. .....
..........
..
......... ..... .
.
.
.
.
.
.
.
...
.....
.........
.......... .....
.............. .....
............... ...
................... ..
.
.
.
.
.
.
.
.
......... ..... ....
............. ....
......... .
...
......... ....
...
.............
.
.
.
.
.
.
.
....
.
....
.....
.........
......
.......
.......
.......
......
.........
..........
...........
...........
.............
.............
............
............
............
................
...............
...............
..............
.................
.................
.................
.................
.................
.................
.................
.................
.................
.................
.................
.............

Rg
R0

tg

Figure 3: The tangency portfolio for R0 < Rg .

What if R0 = Rg ? In this case, 1> V1 (R R0 1) = 0, so the minimum-variance portfolio

z = V1 (R R0 1)
is a self-financing or zero-wealth portfolio of the risky assets. The minimum-variance set is
generated by investing all wealth in the risk-free asset and setting up a long or short position
in portfolio z, i.e., by setting w = z with IR. The corresponding portfolio return is
Rz = (1 1> (z))R0 + R> (z) = R0 + R> z with R> z = C BR0 = /A > 0. Long
positions ( 0) generate the efficient frontier. There is no tangency point it turns out that
the two minimum-variance rays in (, R)-space coincide with the asymptotes of the hyperbola.

The Capital Asset Pricing Model

Next, we want to aggregate investors portfolio choices in order to derive equilibrium properties
of the asset market as a whole. We assume that each investor favours higher expected returns
R and lower variances 2 , and that the compensation R which the agent requires for taking on
another unit of risk 2 increases with the current level of risk 2 . Then, in the standard deviation/mean plane, preferences can be represented by indifference curves of the shape shown in
Figure ??. This guarantees that each investor has a well-defined optimal portfolio, determined
by the point of tangency between the efficient frontier and an indifference curve as shown in
the figure. In particular, each investor demands an efficient portfolio.
The supply side of the asset market is given by a portfolio m of risky investments, the so-called
market portfolio. We assume that all risky assets are in strictly positive supply. The weight of
an asset in the market portfolio is the proportion of the total wealth in the economy which is

Mean-Variance Analysis & CAPM

15

Dr. M. Reisinger

R
....

.
.........
.........
........
.........
........
.
.
.
.
.
.
.
.
..
.........
........
....
........
.....
.........
...
......
.........
.
.
.
.
.
.
.
.
.
.
.
.
.
..
....
......
.........
......
........
....
......
........
.......
.........
....
...... ................
.
.
.
.
.
.
.
.
.
.
...... ........
....
....... .........
..
........ ..........
...
........ ........
.....
....
.......................
.
.
.
.
.
.
.
.
.
.
.
...
.....
...............
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.....
...............
.....
.........
.....
.....
..........
.....
.....
..........
.........
... .....
...
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..... ....
.....
.........
.............
.....
...................
.....
.....................
.....
..............................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
................. ...............
.....
..........................
....
.....
.........
.....
.........
.....
........
.
.
.
.
.
.
.
.
.
.
.
.
....
.. .....
.........
..... ...
..... .....
...

...

....

R0

Figure 4: An optimal portfolio

held in this asset, and hence strictly positive. As all risky assets n = 1, . . . , N are traded, the
market portfolio is directly observable.
Suppose R0 < Rg ; we shall see in a while that this is the only case consistent with equilibrium.
Then, each investor wants to combine borrowing or lending with a long (possibly zero) position
in the tangency portfolio t. In equilibrium, the demand for risky assets must coincide with the
supply of risky assets: t = m, where m is the market portfolio of risky assets. This implies
that the risky-asset market portfolio is efficient (because t is), and the results of the previous
section imply a linear relationship between the expected return on a portfolio and its market
beta (we will discuss this relationship in more detail below).
If we had R0 > Rg , every investor would like to combine some lending with a short (possibly
zero) position in the tangency portfolio. Aggregate demand for risky assets would be given
by t, and market clearing would require t = m. But then 1 = 1> (t) = 1> m = 1, a
contradiction. Finally, if R0 = Rg held true, each agent would want to invest all wealth in the
riskless asset and take a long position in the zero-wealth portfolio z. This would imply that
the total wealth held in the risky assets is zero, which is incompatible with market clearing.
We thus have proven the following result.
Theorem 1 (The Capital Asset Pricing Model) In the presence of a risk-free asset with
return R0 , the expected return on any portfolio p satisfies
Rp = R0 + pm (Rm R0 )
where
Rm R0 > 0.

Mean-Variance Analysis & CAPM

16

Dr. M. Reisinger

This model was derived by Sharpe (1964), Lintner (1965) and Mossin (1965). Both the equation
for Rp and the inequality for Rm reflect the efficiency of the market portfolio. Note that
it is mean-variance analysis which implies two-fund separation and the linear dependence of
expected returns on covariances with efficient portfolios. Equilibrium analysis simply identifies
the market portfolio as being efficient. If the market portfolio were inefficient, or if the risk-free
return were inconsistent with the tangency condition, asset prices (and therefore asset returns)
would have to adjust until equilibrium is attained.
When plotted against market betas, expected portfolio returns lie on a straight line, called the
security market line. It is shown in Figure ??.

Rp
6
.
............
............
............
............
.
.
.
.
.
.
.
.
.
.
.
..
............
............
............
............
............
.
.
.
.
.
.
.
.
.
.
.
...
............
............
............
............
............
.
.
.
.
.
.
.
.
.
.
.
.........
............
............
..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ............................
.......
.
.
.
.
.
.
.
.
.
.
.
..
.
............
..
............
............
............
...
............
.
.
.
.
.
.
.
.
.
.
.
..
............
..
............
.
.
.
.
.
.
.
.
.
.
.
..
............
..
............
.
.
.
.
.
.
.
.
.
.
.
.
............
...
............
.
.
.
.
.
.
.
.
.
.
.
.
..........
..
..............
.
..............
...
..
...
...
...
..

Rm

R0

pm

Figure 5: The security market line in the CAPM

Why should the expected return on a portfolio increase with its market beta? Consider two
portfolios p and q which have identical means and variances of next-period payoffs. Let portfolio
ps payoff be positively correlated with the payoff of the market portfolio, and portfolio qs
payoff negatively correlated. That is, p tends to have high payoffs when the economy is in a
good state, while q tends to have high payoffs when the economy is in a bad state. As one
unit of the payoff is more valuable in the bad state, portfolio q is more desirable than portfolio
p, so p will have a lower price. Since expected returns are calculated as the ratio of expected
payoff over price, portfolio p must have a higher expected return.
We can rewrite the CAPM equation in risk premium or excess return form:
Rp R0 = pm (Rm R0 ).
Thus, the risk premium on a portfolio is beta times the risk premium on the market. This can

Mean-Variance Analysis & CAPM

17

Dr. M. Reisinger

be interpreted in the following way:

Rm R0
m

Rp R0 =
|

{z

risk premium

{z

market price of risk

pm m

| {z }

systematic risk

Only systematic risk, i.e., risk related to market movements, is rewarded in equilibrium. The
market price of this risk is given by the Sharpe ratio of the market portfolio. All other risk does
not earn a reward in form of higher expected returns; the intuition for this is that unsystematic
risk can be diversified away.
In a class exercise, you will discuss the assumptions which we have made to derive the CAPM,
and some non-standard variants of the model.

Appendix
Even a safe government bond is somewhat risky unless it is indexed in a way that eliminates
inflation risk. This appendix is intended for those amongst you who are interested in what
happens to the CAPM when there is no risk-free security. This material is not relevant for the
exam.

Covariances and Expected Returns

In a first step, we consider the covariance properties of the portfolios in the minimum-variance
set when there is no risk-free asset.
The global minimum-variance portfolio possesses the peculiar property that its covariance with
any asset or portfolio is always 1/A. Indeed, for any portfolio p,
>
1
p, R
g ] = p> Vg = p VV 1 = 1 = 2 .
pg = Cov[R
g
A
A

To derive covariance properties of arbitrary portfolios in the minimum-variance set, we go back

to the first order condition
0 = Vw (Rw )1 (Rw )R
for a minimum-variance portfolio w. Suppose w 6= g, and let p be any other portfolio. Premultiplying the first order condition with w> and p> , respectively, we get
2
w
= (Rw ) + (Rw )Rw

and
pw = (Rw ) + (Rw )Rp .
Combining these two equations, we find
2
pw = w
(Rw )(Rw Rp ).

Thus, p has zero covariance with w if and only if its expected return Rp satisfies
2
w
= (Rw )(Rw Rp ).

As minimum-variance portfolios are uniquely determined by their expected return, there is in

particular a unique minimum-variance portfolio z(w) that has zero covariance with w. To locate

Mean-Variance Analysis & CAPM

18

Dr. M. Reisinger

the zero-covariance portfolio z(w) in (, R)-space, we use the fact that a Lagrange multiplier
measures the marginal effect of changing the associated constraint:



1 2 
.
(Rw ) =
 = w
2 R w
R w


This implies
2
w

= (Rw )(Rw Rz(w) ) = w

or


(Rw Rz(w) )
R w

w =


(Rw Rz(w) ).
R w

In other words, the zero-covariance portfolio z(w) can be located as follows in (, R)-space:
draw the tangent to the minimum-variance hyperbola at the point corresponding to w; the
intercept of the tangent determines Rz(w) ; draw a horizontal line through this intercept; the
intersection of this line with the minimum-variance hyperbola determines the standard deviation z(w) . This is illustrated in Figure ??. The diagram shows that if w is efficient, then z(w)
is inefficient, and vice versa.

R
................
.................
.................
................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.....
.................
.................
....
................
.................
.........
.................
.........
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.....
.........
...............
........
..............
......... .............................
.........
..
..............................
.
.
.
.
.
.
.
.
..
......... ............
......................
.......................
............
...........
.
.
.
.
.
.
.
.
.
.
.........
........
...........
.............
....................
.
.
.
.
.
.
.
.
.. .
......... .....
........ .....
........
..
.........
.....
.........
.
.
.
.
.
.
.
...
......
..
........
......................................................................................
......
.......
.......
.......
.......
.........
..........
............
...........
..............
.............
............
...........
.............
................
...............
..............
..............
.................
.................
.................
.................
.................
.................
.................
.................
.................
.................
.................
..........

tg
tz(w)

Figure 6: The zero-covariance portfolio

Now let us go back to the situation where w 6= g belongs to the minimum-variance set and p
is an arbitrary portfolio. We have seen above that
2
pw = w
(Rw )(Rw Rp )

Mean-Variance Analysis & CAPM

19

Dr. M. Reisinger

and
2
0 = w
(Rw )(Rw Rz(w) ).

Eliminating (Rw ), we find the following expression for the expected return on p:
Rp = Rz(w) + pw (Rw Rz(w) )
with pw as defined in the main text. As pw = 0 if and only if pw = 0, the zero-covariance
portfolio z(w) is often termed the zero-beta portfolio associated with w.

The Zero-Beta Capital Asset Pricing Model

Maintaining our assumptions on investor preferences, we can now derive equilibrium properties
of the asset market in the absence of a risk-free security.
As shown in Figure ??, each investor has a well-defined optimal portfolio, determined by the
point of tangency between the minimum-variance hyperbola and an indifference curve.

....
...

...

....

..
.................
.................
..
.................
......
.................
......
.................
.
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
......
....
.......
.................
......
.................
....
......
................
.......
.................
....
......
................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
..
.......
..............
...
.......
.............. ..
........
............... .....
....
........
.............
.....
.....
......... ......................
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.....
......................
.....
......................
.....
.....
....................
.....
.....
..........
.........
... .....
...
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..... ....
.....
............
..............
.....
..................
.....
............. ......
.....
............. ..........
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.................
.....
...
..........................
.....
..
.... .... .....
... .
.
.
.
.
.
.
....
.. .....
.....
..... ...
......
..... .....
.......
.......
.......
.......
.........
..........
...........
...........
.............
.............
............
............
.............
................
...............
..............
..............
.................
.................
.................
.................
.................
.................
.................
.................
.................
.................
.................
...........

....

...

....

Figure 7: An optimal portfolio in the absence of a risk-free asset

Clearly, each investor i demands a portfolio wi on the efficient frontier. To find the overall
demand for assets, we have to weight each of these portfolios by the investors share i of the
total wealth in the economy and add up. This yields
w=

X
i

i wi

Mean-Variance Analysis & CAPM

20

Dr. M. Reisinger

as the aggregate portfolio demanded by the investors. As the minimum-variance set forms a
straight line in IRN , a weighted average of minimum-variance portfolios is again a minimumvariance portfolio. In other words, w is in the minimum-variance set. Moreover, as Rwi > Rg
for all investors i (why?),
!

Rw =

X
i

i Rwi >

i Rg =

i Rg = Rg ,

so w is on the efficient frontier, and w 6= g.

The supply side of the asset market is again given by m, the market portfolio, with all assets
in strictly positive supply.
In equilibrium, supply equals demand: m = w. But this implies that the observed market
portfolio must be mean-variance efficient. By the results of the previous section, this says in
particular that the expected return of a portfolio depends linearly on the covariance or beta
with the market.
Theorem 2 (The Zero-Beta Capital Asset Pricing Model) The expected return on any
portfolio p satisfies
Rp = Rz(m) + pm (Rm Rz(m) )
where
Rm Rz(m) > 0.
This model is due to Black (1972). Note that the second inequality holds because efficiency of
m (Rm > Rg ) implies inefficiency of z(m) (Rz(m) < Rg ). The intuition as to why the expected
return on a portfolio increases with its market beta is exactly the same as in the case with a
risk-free asset.