5 Risk

Which lottery ticket would you pay more for: lottery ticket 1, which pays $1 or $9 with
equal probability, or lottery ticket 2, which pays $4 or $5 with equal probability? Lottery 1
has a 50% chance of high gain, but it is also more risky. Your decision is a question of a
trade-off between risk and potential reward, and it is the subject of this chapter, which has
the following learning objective:
• To present various ways of measuring risk and reward such as the classical mean-variance
theory, and the more recent value-at-risk approach

While utility maximization is a sound theoretical way for making comparisons between
various investment strategies, more specialized and streamlined methods have been used
by practitioners in financial markets. This chapter describes the classic mean-variance
approach of Harry Markowitz for evaluating different portfolios, which is also the basis
for many later developments in the book. The approach is based on comparing the means
and the variances of portfolios. We also discuss a modern way of measuring the risk of a
portfolio, called value-at-risk, or VaR, an approach based on computing probabilities of
large losses. Both these approaches are given in the one-period, static framework, although
there exist extensions to multiperiod models.

5.1 Risk versus Return: Mean-Variance Analysis

The mean-variance analysis was derived by Harry Markowitz in the 1950s as a system-
atic approach to measuring risks in financial markets. He later received a Nobel Prize in
economics for this work. His objective was to establish a criterion for the comparison
of different securities and portfolios as a trade-off between their return—as measured by
expected return—and risk—as measured by the variance of the return. Although in the
original work (Markowitz, 1952) no reference is made to the preferences or utilities of the
investors, we will see that, in his approach, there are several implicit assumptions with
respect to the preferences of the investors or the probability distribution of the returns of
the securities.
The mean-variance approach was originally developed for single-period models. Ac-
cordingly, we assume that the trade is established at time 0 and some random return is
received at time 1. The approach can be extended to multiperiod models, as we discuss
later. Despite the very strong assumptions behind the Markowitz framework, its simplicity
makes it a popular tool for gauging whether a portfolio strategy is “efficient.” We present
this framework next.
154 Chapter 5

5.1.1 Mean and Variance of a Portfolio

If our agent Taf starts with initial capital X (0) and has X (1) at the terminal time 1, the rate
of return is given by

X (1) − X (0)
RX =
X (0)
As usual, we assume that X (1) is a random variable, a result of trading in a market with
assets whose future values are (possibly) random. As we already pointed out, Markowitz’s
framework does not assume any specific distribution for these assets. However, it is assumed
that only the mean and the variance of a portfolio are relevant for evaluating its performance.
There are two possible settings in which that assumption would be fully justified:
• When the distribution of the returns is normal, because a normal distribution is fully
characterized by the mean and the variance.
• When the utility of the investors is quadratic. In a single-period model, the quadratic
utility of an individual who only cares about final wealth X (1) is represented by

E[a X (1) − bX 2 (1)] (5.1)

where a and b are constant. It is clear from this expression that this investor only cares
about the first and the second moment (or, equivalently, the mean and the variance) of the
final wealth.

We therefore assume that the mean and the variance of a given portfolio exist, and we
denote them by
μ = μ X = E[R X ], σ 2 = σ X2 = Var[R X ]

We identify a portfolio strategy with a pair (σ, μ).

Although we take these numbers as given, in practice they have to be computed. There
are two basic approaches. First, using information about past returns of a given security, its
expected value and variance can be estimated using the mean and the variance of a sample
of observations. Second, we can assign probabilities to the possible future returns in our
model, and then compute the (theoretical) mean and the variance. In practice, the former
method is used more frequently, because it makes the parameters of the model consistent
with past returns.
Next, we want to be able to find the mean and the variance of a portfolio with positions in
several assets, if we know the means and the variances of individual assets in the portfolio.
Risk 155

Assume there are N assets in the market, S1 , . . . , S N , with returns

Si (1) − Si (0)
Ri =
Si (0)

their expected values denoted μi , and variances σi2 , i = 1, . . . , N . Denote by i the pro-
portion (weight) of the initial capital X (0) held in asset i. Accordingly,

N
i = 1
i=1
A proportion i can be negative, meaning that Taf is short that asset. Note also that the
number of shares δi of asset i in the portfolio is given by

i X (0)
δi =
Si (0)

The rate of return of the strategy determined by a portfolio  = (1 , . . . ,  N ) is

 N i X (0)
i=1 Si (0) Si (1) − X (0)
R =
X (0)
By dividing all terms in the previous expression by X (0) we get

N
i [Si (1) − Si (0)] 
N
R = = i Ri (5.2)
i=1
Si (0) i=1

which means that the return rate of the portfolio is simply the weighted average of the asset
return rates. Therefore, this statement is also true for the means:
N
μ = i μi
i=1
As for the variance, let us first denote

σi j := Cov[Ri , R j ] := E[(Ri − μi )(R j − μ j )]

the covariance of Ri and R j . We compute the variance for the portfolio in the case N = 2:

σ2 = E[(R − μ )2 ]
= E[{1 R1 + 2 R2 − (1 μ1 + 2 μ2 )}2 ]
 
= E 21 (R1 − μ1 )2 + 21 2 (R1 − μ1 )(R2 − μ2 ) + 22 (R2 − μ2 )2
156 Chapter 5

This produces the following formula for the variance of the return rate for a portfolio
consisting of two assets:

σ2 = 21 σ12 + 22 σ22 + 21 2 σ12 (5.3)

In the Problems section you are asked to show this extension to N assets:

N
σ2 = i  j σi j (5.4)
i, j=1

Here, we denote σii = σi2 .

From this formula we can see how diversifying your portfolio among different assets can
reduce your risk. In particular, assume that there are N independent assets (and, therefore,
with σi j = 0, i = j), all with the same variance σ 2 and the same mean μ. If we hold each
asset with the same proportion 1/N , the portfolio mean is still equal to μ, and its variance
is

1  2
N
σ = σ 2 /N
N 2 i=1

In other words, the variance gets smaller with the number of independent assets in the
portfolio. Recall now that the correlation ρ between two random variables R1 and R2 , with
variances σ12 , σ22 and covariance σ12 , is defined as

σ12
ρ := (5.5)
σ1 σ2

and we have −1 ≤ ρ ≤ 1. If the assets are not completely independent (that is, if their
covariance is different from zero) but they are less than perfectly correlated (the correlation
coefficient is smaller than one in absolute value), then diversification also reduces the
portfolio variance, but less effectively.
When two securities are perfectly correlated and have different means, it is possible to
construct a risk-free portfolio with positive return. We illustrate this point in the following
example.

Example 5.1 (Deriving a Risk-Free Portfolio) Suppose that the stock of company A has an
expected return of 10% and a standard deviation of 8%. The stock of company B has
an expected return of 14% and a standard deviation of 18%. The correlation between A
and B is 1. We want to find the risk-free rate, consistent with no arbitrage. We do so by
constructing a portfolio with zero volatility, hence risk-free, and its return rate has to equal
the risk-free rate. From equations (5.3) and (5.5), we get

σ2 = 2A σ A2 + 2B σ B2 + 2 A  B σ AB = ( A σ A +  B σ B )2

Risk 157

where we have used the fact that ρ AB = 1. We want the variance to be zero, so, since
 B = 1 −  A , we need to have

0 =  A 0.08 + (1 −  A )0.18
with solution  A = 1.8 and, therefore,  B = −0.8. This condition means that for each
$100 of investment, we sell short stock B for a total of $80 and we invest $180 in A. The
return of this portfolio is

μ =  A μ A +  B μ B = 1.8 · 0.1 + (−0.8) · 0.14 = 0.068

The risk-free asset has to have the same return, that is, the risk-free rate has to be 6.8%.

5.1.2 Mean-Variance Efficient Frontier

Suppose now that given two portfolios with the same variance, an investor will prefer the one
with the higher expected return. Different trading strategies will produce different points
(σ, μ) in the plane. We want to get an idea of what kind of mean-variance points we can
achieve by investing in two assets only. Let us represent these two assets by points (σ1 , μ1 )
and (σ2 , μ2 ). Denote the proportion invested in the first asset by , so that the proportion
invested in the second asset is 1 − . For now, we consider only the case when there is
neither borrowing nor short-selling, so 0 ≤  ≤ 1. The mean of the portfolio is

μ = μ1 + (1 − )μ2
The variance will depend on the covariance between the two assets. Recall the correlation
coefficient, defined in equation (5.5). Consider first the case when the two assets are perfectly
correlated, with ρ = 1. In this case we have σ12 = σ1 σ2 . Then the variance of the portfolio
is, by equation (5.3),

σ 2 = 2 σ12 + (1 − )2 σ22 + 2(1 − )σ1 σ2 = [σ1 + (1 − )σ2 ]2

Therefore,

σ = σ1 + (1 − )σ2
This expression means that both the mean and the standard deviation of the portfolio
are linear combinations of the individual assets’ means and standard deviations, respec-
tively. Therefore, by varying the proportion , we will trace a straight line in the (σ, μ)
plane, between the two assets; see figure 5.1. Similarly, if ρ = −1, we can see that (see
Problem 3)

σ = |σ1 − (1 − )σ2 | (5.6)

158 Chapter 5

␮
2

␳  1

␳1

␳  1

␴
Figure 5.1
Mean-variance graph for portfolios of two assets.

This is represented by the two lines connecting at the μ-axis in figure 5.1. For the values of
ρ between −1 and 1, the variance is given by equation (5.3), and varying  corresponds to
a curve between the two assets, such as the one in figure 5.1.
More generally, with N assets, one can obtain portfolios in the region to the right of a
curve such as the one in figure 5.2. This region is called the feasible region. For a fixed
mean return μ, the standard deviation σ for which the point (σ, μ) is on this curve is the
smallest standard deviation that can be obtained by trading in these assets, because the
point on the curve is the leftmost point in the feasible region with that value of μ. The very
leftmost point on the curve is called the minimum variance portfolio, because it represents
the portfolio that has the smallest possible variance. The part of the curve below this point
is not interesting—for each point on the lower part of the curve there is a point on the upper
part that has the same variance but a higher mean. The upper part of the curve is called
the efficient frontier, since our investor Taf prefers smaller variance and higher mean for a
portfolio:
The points on the efficient frontier are the points that correspond to the highest mean for
a given level of variance, as well as the points that correspond to the lowest variance for a
given mean level.
If one of the available assets is a risk-free asset (therefore, with zero variance) and has a
(nonrandom) return R, then the feasible region looks like the one in figure 5.3. In order to
see that fact, note that a combination of an asset (μ1 , σ1 ) with the risk-free asset results in
Risk 159

Efficient frontier

Feasible region

Minimum variance point

Figure 5.2
The efficient frontier is the upper part of the curve.

Figure 5.3
Efficient frontier with risk-free asset: The efficient frontier is the upper line. Point P is the mutual fund of risky
assets that can be used for creating efficient portfolios of risky assets and the risk-free asset.
160 Chapter 5

the mean and standard deviation

μ = R + (1 − )μ1 , σ = (1 − )σ1

(Recall that any nonrandom quantity has zero variance and zero covariance with any other
quantity.) So, when we invest in the risk-free security and any portfolio on the efficient
frontier, we can attain any combination (σ, μ) on the straight line between the portfolio
and the risk-free rate of return R on the y-axis by changing the proportion . The efficient
frontier is now the upper line of the feasible region. The curve inside the feasible region in
figure 5.3 is the boundary of the feasible region when there is no risk-free asset. The point
P where this curve is tangent to the efficient frontier line corresponds to a special portfolio
(or mutual fund):
Any portfolio on the efficient frontier line can be obtained by trading in the fund P and
the risk-free asset.
This result is called the one-fund theorem. It holds because any point on a line can be
obtained as a linear combination of the two points R and P on the line. It means that, for
the mean-variance investors, in order to be optimal it is sufficient to invest in the single
mutual fund P and the risk-free asset. Among all the points on the efficient frontier without
the risk-free asset, P is the most important point, because it allows the investor to attain
the portfolios on the line that is the efficient frontier in the presence of the risk-free asset.
We also observe that the points on this line to the right of and above P can be attained by
short-selling the risk-free security and investing the proceeds in the portfolio P.

5.1.3 Computing the Optimal Mean-Variance Portfolio

The mean-variance optimization problem postulates that we consider only portfolios with
a given mean return rate equal to μ. Among those, we want to find the portfolio that results
in the smallest variance. In other words, with N assets and portfolio weights denoted by
i , i = 1, . . . , N , we need to solve the optimization problem that consists in minimizing
the variance of the portfolio,

N
1
2
min i  j σi j
i, j=1

subject to the constraints


N 
N
i μi = μ, i = 1 (5.7)
i=1 i=1

The factor 12 in the optimization problem is introduced for convenience only: it simplifies
some of the resulting expressions, and, obviously, minimizing the value of half the variance
Risk 161

is equivalent to minimizing the variance. Since this is an optimization problem with two
constraints, we use two Lagrange multipliers λ1 and λ2 , and form a Lagrangian:
 N   N 
N  
L=2 1
i  j σi j − λ1 i μi − μ − λ2 i − 1
i, j=1 i=1 i=1

Then we differentiate the Lagrangian with respect to all i and set derivatives equal to
zero. This step would give us the following N equations for the optimal mean-variance
weights:

N
σi j  j − λ1 μi − λ2 = 0, i = 1, . . . , N (5.8)
j=1

However, we have N + 2 unknowns, i ’s, and λi ’s. The remaining two equations are the
constraint equations (5.7).
We show equations (5.8) for the case N = 2:


L = 12 12 σ12 + 22 σ22 + 21 2 σ12 − λ1 (1 μ1 + 2 μ2 − μ) − λ2 (1 + 2 − 1)

Setting the derivative of L with respect to 1 equal to zero, we get

1 σ12 + 2 σ12 − λ1 μ1 − λ2 = 0
Similarly for the derivative with respect to 2 . This result is in agreement with equation (5.8)
when N = 2.

Example 5.2 (Two Assets) In the case of only two assets, the problem becomes degenerate,
in the sense that the portfolio weights are determined directly from the constraints (5.7).
Indeed, suppose that we want to achieve a mean μ with two assets. Then, denoting by 
the proportion in asset 1 [and, therefore, (1 − ) the proportion in asset 2], we need to have

μ1 + (1 − )μ2 = μ
This gives
μ − μ2
=
μ1 − μ 2

Suppose for example that μ1 = 0.15 and μ2 = 0.10. Then we have

1 = 20(μ − 0.1), 2 = 3 − 20μ

The portfolio variance is given by

σ 2 = 400(μ − 0.1)2 σ12 + (3 − 20μ)2 σ22 + 40(μ − 0.1)(3 − 20μ)σ12
162 Chapter 5

This equation determines the mean-variance curve in the (σ, μ) plane. Suppose, for sim-
plicity, that the assets are uncorrelated, σ12 = 0, and that σ12 = σ22 = 0.01. The variance is
then
σ 2 = 4(μ − 0.1)2 + (0.3 − 2μ)2

Let us find the minimum-variance portfolio. We look for an expected return μ such that
the portfolio with such a return has the lowest possible variance. Setting the derivative with
respect to μ equal to zero, we obtain
μ = 0.125

The variance corresponding to this mean is

σ 2 = 0.005

and it is the minimum variance for the portfolios consisting of these two assets. The pro-
portions to be held in the two assets are
1 = 20(0.125 − 0.1) = 0.5 = 50%, 2 = 50%

Example 5.3 (Three Assets) Consider now the case of three uncorrelated assets, all having
the same variance equal to 0.01. Suppose also that
μ1 = 0.1, μ2 = 0.2, μ3 = 0.3

Then, equations (5.8) become

0.011 − 0.1λ1 − λ2 = 0
0.012 − 0.2λ1 − λ2 = 0
0.013 − 0.3λ1 − λ2 = 0
Together with the constraints
0.11 + 0.22 + 0.33 = μ, 1 + 2 + 3 = 1

these equations can be solved for i , i = 1, 2, 3, in terms of μ (see Problem 9).

To recap, equations (5.8) give us a way to find optimal weights for the assets in our
portfolio (that is, the weights that yield the lowest possible variance for a given expected
return) in a single-period model, assuming we know the variance-covariance structure, as
well as the expected return of the assets. We will use mean-variance optimization later, in
our discussion of equilibrium in chapter 13.
Risk 163

5.1.4 Computing the Optimal Mutual Fund

It may be useful for practical purposes to identify the mutual fund P from the one-fund
theorem. We first point out that the first-order condition (5.8) is satisfied regardless of the
pool of assets. That is, even if one of the assets is risk-free with zero variance and zero
covariance with all other assets, equation (5.8) has to hold. Then the straight line that goes
through R and P is also characterized by equation (5.8). Out of all the points on that line,
we look for the point P that implies zero investment in the risk-free asset.
More precisely, suppose that there is a risk-free asset with mean return μ1 = R, and that
Taf wants to find the proportions i , i = 2, . . . , N of the risky assets held in the fund P.
This fund is on the efficient frontier (which is the straight line that goes through R and P),
and hence, the proportions i can be found from equation (5.8), taking into account that
the proportion 1 of the risk-free asset in the fund is zero. In other words, we have to solve
the mean-variance optimization problem with

N
1 = 0, i = 1
i=2

Since σ1 j = 0, j = 1, . . . , N , from the equation with i = 1 in equation (5.8) we get

λ2 = −λ1 R
Substituting this in equations (5.8) with i = 2, . . . , N , we get the system

N
σi j  j = λ1 (μi − R), i = 2, . . . , N (5.9)
j=2

We can solve this system for i ’s in terms of λ1 , and then find λ1 from the condition
N
i=2 i = 1.

Example 5.4 (Optimal Mutual Fund) Taf can invest in asset 1 with μ1 = 0.2, σ1 = 0.4
and asset 2 with μ2 = 0.3, σ2 = 0.5, with correlation ρ = 0.2, and in the risk-free asset
with return R = 0.05. He wants to identify the weights of asset 1 and asset 2 in the mutual
fund P. The system (5.9) becomes

σ12 1 + σ12 2 = λ1 (μ1 − R)

σ22 2 + σ21 1 = λ1 (μ2 − R)
Since σ12 = σ21 = ρσ1 σ2 = 0.04, we get
0.161 + 0.042 = 0.15λ1
0.252 + 0.041 = 0.25λ1
164 Chapter 5

Solving this system we get

1 = 0.7162λ1 , 2 = 0.8854λ1

From 1 + 2 = 1, we get λ1 = 0.6244 and

1 = 0.4472, 2 = 0.5528

This result means that first Taf should decide how much money to invest in the risk-free
asset and how much in the risky assets, and the money in the risky assets should be split
into about 45% in asset 1 and 56% in asset 2.
5.1.5 Mean-Variance Optimization in Continuous Time∗

The mean-variance optimization problem is usually considered in the single-period frame-

work. However, using the duality/martingale approach to portfolio optimization, we can
analyze the problem in continuous time, too. For simplicity, we adopt the Black-Scholes
model with one stock with price S and one risk-free security with price B, that is,

d S(t)
= μ dt + σ d W (t)
S(t)
d B(t)
= r dt
B(t)
with μ, σ , and r constant. In this setting, our investor Taf tries to minimize the variance of
the value of the discounted final wealth X¯ (T ):

min E[ X̄ 2 (T )] − (E[ X̄ (T )])2 (5.10)

under the constraint on the mean

E[ X̄ (T )] = m (5.11)

and the usual budget constraint

E[Z (T ) X̄ (T )] = x (5.12)

where x is Taf’s initial wealth and Z is the risk-neutral density process discussed in chapter 4.
We formulate the problem with the discounted wealth for convenience only, and without
loss of generality. We have to assume that the market price of risk θ is different from zero,
so that Z (T ) is different from one.
We could solve the problem using the rigorous duality approach as in utility maximization
sections, but we choose to present the more intuitive Lagrange-multipliers approach. We
Risk 165

write the Lagrangian as

E[X̄ 2 (T )] − m 2 − 2λ1 {E[X̄ (T )] − m} − 2λ2 {E[Z (T )X¯ (T )] − x}

The factor 2 in front of the multipliers λi is introduced for convenience only. Proceeding in
a heuristic fashion, we take a derivative with respect to X̄ (T ) and set it equal to zero to get
the following expression for the discounted optimal terminal wealth:

X̄ (T ) = λ1 + λ2 Z (T ) (5.13)

In particular, since Z is a martingale with expectation one, taking expectations we calculate

the mean as
E[X̄ (T )] = λ1 + λ2

which, using equation (5.11), gives us

λ2 = m − λ1 (5.14)

As for the budget constraint, we multiply all terms of equation (5.13) by Z (T ) and take
expectations. Using again the fact that Z (T ) is a martingale with initial value 1, we get

E[Z (T )X¯ (T )] = λ1 + λ2 E[Z 2 (T )]

From equation (5.12) this expression has to be equal to x, and, also using equation (5.14),
we deduce

x − m E[Z 2 (T )] m−x
λ1 = , λ2 = (5.15)
1 − E[Z 2 (T )] 1 − E[Z 2 (T )]

Since
Z 2 (T ) = e−2θ W (T )− 2 θ T eθ = M(T )eθ
4 2 2 2
T T

where

M(t) := e−2θ W (t)− 2 θ

4 2
t

is a martingale process, we get

E t [Z 2 (T )] = M(t)eθ
2
E t [Z (T )] = Z (t), T
(5.16)

and, since M(t) is a martingale with M(0) = 1,

E[Z 2 (T )] = eθ
2
T
(5.17)

Substituting in equations (5.15), we get the values of the multipliers.

166 Chapter 5

Next, we want to compute the optimal portfolio process. We recall that Z X̄ is a martingale
process, so that, using equation (5.13),

Z (t)X¯ (t) = E t [Z (T )X¯ (T )] = E t [λ1 Z (T ) + λ2 Z 2 (T )] (5.18)

Hence, using equations (5.16) in equation (5.18),

Z (t) X̄ (t) = λ1 Z (t) + λ2 eθ T M(t)

2
(5.19)

Using Itô’s rule, we get

d[Z (t) X̄ (t)] = −λ1 θ Z (t) d W (t) − 2λ2 eθ T θ M(t) d W (t)

2
(5.20)

However, we recall that

d[Z (t) X̄ (t)] = Z̄ (t)[π(t)σ − θ X (t)] d W (t) (5.21)

where π is the amount held in the stock. We can now compute the optimal portfolio π
matching the right-hand sides of equations (5.20) and (5.21):

Z̄ (t)[π(t)σ − θ X (t)] = −λ1 θ Z (t) − 2λ2 eθ T θ M(t)

2

This, along with equation (5.19), implies

Z̄ (t)π(t)σ = −λ2 eθ T θ M(t)

2
(5.22)

where λ2 is given by equations (5.15) and (5.17), which depend on m, so that the amount
to be invested in the risky security is a function of the desired expected value of wealth. In
particular, at initial time t = 0 when Z̄ (0) = M(0) = 1, we get

π(0)σ = −λ2 eθ T θ
2
(5.23)

Finally, let us compute the equation of the efficient frontier in this setup. The variance of
the discounted terminal wealth can be calculated as
E[X̄ 2 (T )] − m 2 = E[X̄ (T ){λ1 + λ2 Z (T )}] − m 2 = λ1 m + λ2 x − m 2

where the first equality follows from equation (5.13) and the second equality uses the
constraint (5.12). Substituting for the values of λi , we can calculate this expression as
(m − x)2
Var[ X̄ (T )] =
E[Z 2 (T )] − 1
This expression means that the mean-variance frontier is described by the linear relationship
|m − x|
σ X̄ (T ) = 
E[Z 2 (T )] − 1
Risk 167

We make the obvious observation that, in order to achieve the mean discounted wealth
equal to the initial capital, that is, if m = x, the investor should put all the money in the
risk-free security.

5.2 VaR: Value at Risk

A company doing business in world markets is exposed to all kinds of market risks: risks of
fluctuating foreign exchange rates, of fluctuating stock prices and interest rates, of fluctuating
commodity prices, of fluctuating correlations between various market variables, and so on.
For the purposes of risk management by company executives or supervision by regulatory
agencies, it is convenient to have simple measures of risk that can be easily understood
and interpreted. In particular, it is convenient to be able to come up with only a single
number as a measure of risk. It has become the industry standard to use, for this purpose, a
number related to the probability of large losses. This is called the value-at-risk approach,
or VaR approach (not to be confused with Var[·] as in “variance”). For example, a VaR of
an investment company that is too high is an indication that the company’s positions are
too risky and have to be modified. Like most other successful approaches or models, this
too is based on a simplification of reality and has a number of shortcomings. Nevertheless,
it is a benchmark method and a standard for measuring risk, from which other, more
complex and more sophisticated approaches can be developed and to which they may be
compared.

5.2.1 Definition of VaR

Suppose we are interested in large losses of a given portfolio and, specifically, in those losses
that are likely to happen not more often than once in hundred trading days. We say that
we are interested in the 99% confidence level and in a daily value at risk. For example, a
company might report its daily VaR at the 99% level to be $1 million. This statement means
that the company estimates that there is less than a 1% chance that the company will lose
$1 million during the next day. We have the following exact definition:

Daily value at risk at a 99% confidence level is the smallest number x for which the
probability that the next day’s portfolio loss will exceed x is not more than 1%.

In mathematical notation, if we denote by L the loss of the next day, we have

P[L ≥ x] ≤ 0.01 (5.24)

and VaR is the smallest number x for which condition (5.24) holds. See figure 5.4 for a
graphical illustration, when we assume a normal distribution (we discuss this assumption
later).
168 Chapter 5

Loss density

2.33
6 4 2 0 2 4 6

Figure 5.4
Value at risk: There is 1% probability (area) that the loss will be larger than the value 2.33 on the x-axis.

We can define VaR for other time periods. For regulatory purposes, a 10-day VaR is
typically used. For example, if we say that a 10-day VaR is $20 million, at a 95% level, we
mean that there is less than 5% chance our portfolio will lose more than $20 million in the
next 10 days.
For concreteness, from now on we consider only the daily VaR at the 99% level.
5.2.2 Computing VaR

How do we compute the VaR of a portfolio? The most difficult thing about this question is
deciding what kind of distribution to assume for the portfolio returns. Are they normally
distributed? Could there be jumps in the returns? Should we believe past experience? There
are several approaches to this problem:

Historical, Nonparametric VaR In the historical, nonparametric approach, we look at

the sample of past returns, and we find the historical loss size for which there were only
1% of days with larger losses. This gives us the estimated VaR of the portfolio. Even if we
believe that the past is a good estimate of the future in a particular situation, there are still
difficulties in this approach: How large a sample (how many days in the past) should we
choose? Should we discard unusual days with extremely high losses, say, during a market
crash? Should we count only the trading days, or all days? How do we deal with a portfolio
that is dynamic and has changed a lot with regard to the positions held in various assets? In
short, this is a simple approach conceptually, but often tricky to successfully implement in
practice.

Model-Based, Parametric VaR The model-based, parametric approach is more fre-

quently used. We assume that the portfolio returns have a particular distribution, and we
compute the VaR using probabilistic methods. There are two main tasks: we have to choose
Risk 169

which distribution the returns come from, and we have to estimate the parameters of that
distribution, using historical data or otherwise.
One of the easiest solutions, though not always very realistic, is to assume that the rates
of return (sometimes it is more realistic to use log-returns) have a normal distribution with
mean μ and variance σ 2 . The parameters μ and σ 2 have to be estimated. Denoting by R
the rate of return and by X (t) the portfolio value at time t, we have

X (1) − X (0)
R=
X (0)
so that the actual loss {which is the negative of the profit, that is −[X (1) − X (0)]} is

L = −X (0)R

For example, if the portfolio returns −10% on the initial investment of $100,000, the loss
is $10,000.
When R is normally distributed, the variable Z = (R − μ)/σ has the standard normal
distribution, and we can compute the VaR from the following equation:

0.01 = P[L ≥ x] = P[−X (0)R ≥ x] = P[X (0)R ≤ −x] = P[R ≤ −x/ X (0)]
   
R−μ −x/ X (0) − μ −x/ X (0) − μ
=P ≤ =P Z≤ (5.25)
σ σ σ

Using a standard normal distribution table (Z -table) or a software package such as Excel,
we can find that

P[Z ≤ −2.33] = 0.01

Therefore, we get the VaR from

−x/ X (0) − μ
= −2.33
σ
resulting in the formula for 99% VaR with normal distribution:

VaR = X (0)[2.33σ − μ] (5.26)

For example, if a portfolio has a mean daily return of 0.03% and a standard deviation of
1%, and its current value is $1 million, then VaR = 0.023X (0) = $23,000.

Excel Tip: In order to find x for which P[Z ≤ x] = y, use the command
NORMSINV(y)
170 Chapter 5

5.2.3 VaR of a Portfolio of Assets

If a portfolio is changing its positions in different assets in a dynamic fashion, it may not be
appropriate to model the returns of the portfolio directly as coming from a fixed distribution.
Rather, it makes more sense to model the returns of the individual assets in the portfolio
and then to derive conclusions for the whole portfolio. As an example, consider a portfolio
of two stocks, with δ1 shares of stock S1 and δ2 shares of stock S2 :

X = δ1 S1 + δ2 S2

Denote by i the weight of stock i in the portfolio:

δi Si
i =
X
Recall that the rate of the return R of the portfolio is the weighted sum of the rates of return
of the assets in the portfolio:
R = 1 R1 + 2 R2
If we assume that the stocks’ return rates are normally distributed with means μi , variances
σi2 , and correlation ρ, then the return rate R of the portfolio is also normally distributed
with mean

μ = 1 μ1 + 2 μ2
and variance

σ 2 = 12 σ12 + 22 σ22 + 21 2 σ1 σ2 ρ

With this knowledge, we can compute the VaR of the portfolio using equation (5.26). See
Problems 18 and 19 for examples.
We mentioned that the normal distribution is easier to use than other distributions. This
property is not so much true for one-dimensional distributions, but it becomes more rele-
vant when dealing with a joint distribution of many random variables. One of the biggest
difficulties in computing VaR is estimating the correlations between the different assets in
a portfolio. Everything becomes simpler if we assume that the assets’ return rates follow a
multivariate normal distribution. Theoretically, the normal-distribution assumption is jus-
tified by the central limit theorem if the portfolio is well diversified, that is, if it consists
of many assets that are close to being independent. The central limit theorem says that a
sum of a large number of independent random variables behaves approximately as a normal
random variable (see chapter 16).
Risk 171

5.2.4 Alternatives to VaR

The VaR approach is widely accepted and easy to understand. However, it has some dis-
advantages. One disadvantage of VaR is that, if a large loss happens, VaR tells us nothing
about the actual size of the loss. Moreover, when calculating VaR, it is usually assumed that
normal market conditions prevail, disregarding possible market crashes or extreme moves of
market variables. And VaR, being a single number, does not indicate which of the portfolio
components is responsible for the largest risk exposure. For these reasons, other methods for
risk measurement have been developed, such as stress testing and scenario analysis. These
methods typically involve many simulations of possible changes in variables affecting the
portfolio, such as prices, interest rates, and correlations. For every simulation the profit or
loss is calculated, giving us an idea of how the portfolio would behave in adverse market
conditions. These simulations can be random, or they can be based on market moves from
the past when they were of extreme size, such as during market crashes.
Another problem with VaR is that the VaR of a combination of two positions may be
larger than the sum of the VaRs of the individual positions. This possibility goes against
the usual diversification feature, that the risk of a diversified portfolio is no larger than the
combined risk of the portfolio components.
In the next section we mention a real-world example that shows a need for risk
management.

5.2.5 The Story of Long-Term Capital Management

Perhaps the best known recent example of how things can go wrong even when sophisti-
cated modeling is used is the example of the hedge fund Long-Term Capital Management
(LTCM) and its near demise in September 1998. The principals of the fund included Wall
Street star traders, as well as two scholars whose names we encounter a lot in this book:
Robert C. Merton and Myron Scholes. The fund lost more than 50% of its value (mea-
sured in hundreds of billions of dollars) during the crisis triggered by the devaluation of the
Russian ruble and Russia’s freeze on its debt payments. A part of the LTCM strategy was
to anticipate that the spreads between different rates would become narrower. In particular,
the spreads between the rates of treasury bonds and corporate bonds were at an unusually
high historical level. Similarly, it was expected that the spreads between the interest rates
of several Western European countries would become smaller, as a result of their becoming
part of the European Union. However, the Russian crisis made investors “fly to safety”
and become very conservative. This trend forced the spreads to go the other direction and
become wider. Like most other hedge funds, LTCM was highly leveraged, meaning it had to
use a lot of borrowed funds to sustain its positions. Once the crisis started, the fund started
receiving margin calls, forcing it to try to sell some of its assets. Unfortunately, there was a
172 Chapter 5

sudden drop in liquidity, and it was difficult for LTCM to unwind positions (sell portfolio
assets) at fair market prices. Even worse, many market participants were copying the trades
of the previously successful LTCM, thus amplifying price movements. Eventually, the fund
was bailed out by the concerted effort of a number of financial institutions that provided
enough cash for its survival. The effort was initiated by the U.S. government, through me-
diation led by the Federal Reserve Bank of New York. The motivation for this effort was to
help stabilize the markets, because a bankruptcy of such a big fund as LTCM would have
had a major destabilizing effect.
The LTCM fund had its risk management system, including VaR and scenario testing.
However, it seems that the testing was done in conditions that did not include market moves
as extreme as the ones that actually occurred, partly because such moves had never before
been experienced.

Summary

The mean-variance theory compares portfolio strategies on the basis of the mean and the
variance of their return. The strategies can be represented as points in the mean-variance
plane. The strategies with the lowest variance for a given level of the mean return form
the efficient frontier in the plane. In order to have an efficient strategy composed of holdings
in risky assets and a risk-free asset, it is sufficient to invest in one mutual fund of risky
assets and in the risk-free asset. Optimal mean-variance strategies can be computed in
various models.
A modern way of measuring the risk of a portfolio is using value at risk, or VaR. It
is related to the probability of large losses. There are different models in which we can
compute VaR. Even if a company calculates the VaR of its positions and other measures of
risk on a regular basis, it can still run into big losses if the models are not adequate.

Problems

1. Consider two investors who care only about the means and variances of their investments.
Investor A is indifferent between portfolio 1 with expected return of 10% and standard
deviation of 15% and portfolio 2 with expected return of 18% and standard deviation of
20%. Investor B is indifferent between portfolio 3 with expected return of 12% and portfolio
4 with expected return of 15%, where the standard deviation of portfolios 3 and 4 are the
same as of portfolios 1 and 2, that is, equal to 15% and 20%, respectively. Which of the two
investors would you say is more risk-averse?
 Risk 173

2. Prove equation (5.4).

3. Prove equation (5.6) for σ in terms of  in the case of two assets with correlation
ρ = −1, means μ1 and μ2 , and standard deviations σ1 and σ2 . What does this equation
look like in terms of μ rather than in terms of ?
†4. Consider a mutual fund F that invests 50% in the risk-free security and 50% in stock A,
which has expected return and standard deviation of 10% and 12%, respectively. The
risk-free rate is 5%. You borrow the risk-free asset and invest in F so as to get an expected
return of 15%. What is the standard deviation of your investment?
5. Suppose that securities A and B are perfectly and positively correlated, meaning that
R A = a + b R B for some constants a and b > 0, with expected returns 8% and 12%,
respectively, and standard deviations 15% and 25%, respectively. Compute the risk-free
rate (or the rate of return of a risk-free portfolio).
6. You can invest in asset 1 with μ1 = 0.1, σ1 = 0.3 and asset 2 with μ2 = 0.2, σ2 = 0.5,
with correlation ρ = 0.2. Find the minimum variance point and the portfolio strategy that
attains it.
7. Repeat the previous problem if a risk-free asset with return R = 0.05 is also available.
8. Find the portfolio weights in the optimal mutual fund P for the two risky assets from
the two previous problems.
9. Complete the solution to Example 5.3.
†10. You can invest in asset 1 with μ1 = 0.1, σ1 = 0.3 and asset 2 with μ2 = 0.2, σ2 = 0.5,
with correlation ρ = 0.2. You can also invest in the risk-free asset with return R = 0.05.
Find the optimal mean-variance portfolio for the given mean return μ = 0.2.
11. Consider three securities A, B, and C with expected returns of 10%, 15%, and 12%,
respectively, and standard deviations of 8%, 18%, and 20%. The correlation coefficients are
0.7 between A and B, 0.1 between A and C, and 0.4 between B and C. What is the portfolio
on the efficient frontier that corresponds to an expected return of 12%? Note: You can
provide either an exact answer to this question or an approximate answer. It is acceptable to
find the variance for a number of portfolios (say 50) and choose the one with the smallest
variance. (You can use a software program for this purpose.)
†12. Suppose that the risk-free rate is 5%. There are three risky portfolios A, B, and C with
expected returns 15%, 20%, and 25%, respectively, and standard deviations 5%, 10%, and
14%. You can invest in the risk-free security and only one of the risky portfolios. Which
one of them would you choose? What if the risk-free rate is 10%?
 174 Chapter 5

*13. Consider a Black-Scholes model with μ = 0.1, σ = 0.3, r = 0, and a mean-variance

investor starting with initial capital x = 1 who wants to attain the mean wealth level of m
in T = 1 year. Draw a graph of the optimal portfolio amounts π(0) to be held in the stock
today for the values of m = 1, 1.1, . . . , 1.9, 2.
†14. Compute the historical daily 90% VaR of a portfolio whose daily losses in the last
10 days were, in millions of dollars (minus sign indicates a profit):

1, −0.5, −0.1, 0.7, 0.2, 0.1, −0.2, −0.8, −0.3, 0.5

15. Compute the historical daily 80% VaR of a portfolio whose daily losses in the last
five days were, in millions of dollars (minus sign indicates a profit):
0.3, −0.4, 0.1, −0.2, 0.2
†16. Compute the daily 99% and 95% VaR of a portfolio whose daily return is normally
distributed with a mean of 1% and a standard deviation of 0.5%. The current value of the
portfolio is $1 million.
17. Compute the daily 99% and 95% VaR of a portfolio whose daily profit or loss size
is normally distributed with a mean of $10,000 and a standard deviation of $5,000. The
current value of the portfolio is $2 million.
†18. Small investor Taf has 70% of his portfolio invested in a major market-index fund and
30% in a small-stocks fund. The mean monthly return rate of the market-index fund is
1.5%, with standard deviation 0.9%. The small-stocks fund has the mean monthly return
rate of 2.2% with standard deviation of 1.2%. The correlation between the two funds is 0.13.
Assume normal distribution for the return rates. What is the monthly VaR at 99% level for
Taf’s portfolio if the portfolio value today is $100,000?
19. Your portfolio is made up like this: you hold 60% in the U.S. stock market, with a mean
monthly return rate of 1% and standard deviation 2%; you hold 10% in the global stock
market, with a mean monthly return rate of 0.6% and standard deviation 1%; you hold 30%
in risk-free assets with annual interest rate 5%. Assume normal distribution for the return
rates of the stock markets. The correlation between the U.S. market and the global market
is 0.25. What is the monthly VaR at a 99% level for your portfolio if the portfolio value
today is $100,000?
*20. Choose a company listed on the NYSE and collect the daily stock prices for the last
100 days. (Alternatively, simulate 100 stock prices from some distribution.)
a. Compute the daily historical VaR.
b. Compute the daily parametric VaR, assuming the normal distribution for the relative
returns.
Risk 175

Further Readings

The original mean-variance approach is due to Markowitz (1952). A general advanced

exposition can be found in LeRoy and Werner (2001). The continuous-time treatment of the
mean-variance frontier is taken from Cvitanić, Lazrak, and Wang (2003). Nice treatments
of the value-at-risk approach are Duffie and Pan (1997) and Jorion (1997). An application
of VaR to optimal portfolios is in Basak and Shapiro (2001). More general treatment for
loss distributions can be found in Embrechts, Kluppelberg, and Mikosch (1997).