Professional Documents
Culture Documents
probability distribution. Then X1 and X2 have the same expected value m = E{Xi } and
variance σ2 = Var{Xi }.
Ann and David decide to divide the total income into equal shares, so each will have the
random income
1
Y = (X1 + X2 ).
2
Then
1
E{Y } = (m + m) = m,
2
that is, the same as before sharing the risk. On the other hand, the variance
1 2 σ2
Var{Y } = (σ + σ2 ) = ,
4 2
is half as large.
Now, consider n participants of a mutual risk exchange, and denote their random incomes
by X1 , ..., Xn . Assume the X’s to be i.i.d., m = E{Xi } and σ2 = Var{Xi }.
Then the income for each is
X1 + ... + Xn
Y= .
n
σ2
E{Y } = m, while Var{Y } = ,
n
and for large n, the variance is close to zero.
Also,
X1 + ... + Xn
P → m = 1.
n
We also right the same as
X1 + ... + Xn a.s.
→ m.
n
Thus, for a large number of participants, the risk of each separate participant
may be reduced nearly to zero.
1
2 1. COMPARISON OF RANDOM VARIABLES
X = {X}
and define a rule of comparison of these r.v.; that is, for each pair (X,Y ) of r.v.’s, we
determine whether X is better than Y , or X is worse than Y , or these two random variables
are equivalent for us.
We may also say that X is not worse than Y , etc.
Formally, we write this as
X % Y,
X ≺ Y,
Y ≻ X,
If simultaneously X % Y and Y % X, we write X ≃ Y .
The monotonicity property:
EXAMPLE 1. Let two r.v.’s, X = X(ω) and Y = Y (ω), be defined on a sample space Ω
consisting of only two outcomes: ω1 and ω2 , and
X(ω1 ) = 1, X(ω2 ) = 3,
Y (ω1 ) = 1, Y (ω2 ) = 2.
Let V (X) be a function taking on numerical values. We say that an order % is preserved,
or completely characterized, by V (X) if for any X,Y ,
X % Y ⇔ V (X) ≥ V (Y ).
X % Y ⇔ E{X} ≥ E{Y }.
In this case, V (X) = E{X}.
Note that from the mean-value criterion’s point of view, for example, r.v.’s
100 with probability 1/2
X= , and Y = 50 (1.2.1)
0 with probability 1/2
are equivalent.
1. A General Framework and First Criteria 5
where the minus reflects the fact that the quality decreases as the variance increases.
The positive parameter τ is usually called a tolerance to risk.
Note also that often, instead of (1.2.2), people consider V (X) = τmX − σ2X , but the differ-
ence is non-essential.
The function V (X) in (1.2.2) preserves the corresponding preference order % among
r.v.’s:
X % Y ⇔ τmX − σX ≥ τmY − σY .
The mean-variance criteria are very popular, especially in Finance, and at first glance
look quite natural. However, there are situations where the choice of such criteria may
contradict common sense.
EXAMPLE 1. Let X = 0, a number a ≥ 1, and
a with probability 1a ,
Y=
0 with probability 1 − 1a .
Clearly,
E{Y } = 1, Var{Y } = E{Y 2 } − (E{Y })2 = a − 1.
Then,
√ √
V (Y ) = τ · 1 − a − 1 = τ − a − 1, while V (X) = τ · 0 − 0 = 0.
So, whatever τ is, we can choose a sufficiently large a for which V (Y ) < 0. On the
other hand, V (X) = 0, and under the mean-variance criterion, Y is worse than X, whereas
P(X ≤ Y ) = 1.
EXAMPLE 2. Let X take on values from [1, ∞), and P(X > x) = 1/xα for all x ≥ 1 and
some α > 2. This is a version of the Pareto distribution.
α α
mX = , and σ2X = .
α−1 (α − 2)(α − 1)2
One may observe the same phenomenon for V (X) equal to almost
any function g(mX , σX ) of the mean and standard deviation.
Moreover, for any r.v. X, we may point out a r.v. Y such that P(Y ≥ X) = 1
whereas V (Y ) < V (X).
Nevertheless, it is worth emphasizing that the reasoning above does not mean
that we should not use mean-variance criteria, but it does mean that we should
be cautious .
1. A General Framework and First Criteria 7
0 qγ x 0 qγ x 0 qγ x 0 qγ x x
qγ =0
We come to the criterion under discussion. Another term in use is the capital-at-risk
criterion.
Let γ be a fixed level of probability, viewed as sufficiently small. Assume that an investor
does not take into consideration events whose probabilities are less than γ. Then, for such
an investor the worst, smallest conceivable level of the income is qγ .
Let, for instance, γ = 0.05. Then q0.05 is the smallest value of the income among all
values which may occur with 95% probability. One may say that q0.05 is the value at 5%
risk. Note that qγ may be negative, which corresponds to losses.
The VaR criterion is defined as
X % Y ⇔ qγ (X) ≥ qγ (Y ),
EXAMPLE 1. Let a r.v. X (say, a random income) take on values 0, 10 with proba-
bilities 0.1 and 0.9, respectively, and let a r.v. Y take on the same values with respective
probabilities 0.07 and 0.93. The reader is suggested to check that for γ = 0.05, we have
qγ (X) = qγ (Y ) = 0 (look, for example at Fig.1d). So, X and Y are equivalent under the VaR
criterion. However, for γ = 0.08 we have qγ (X) = 0 while qγ (Y ) = 10, that is, X is worse
than Y . So, the result of comparison depends on γ.
EXAMPLE 2. Let X be normal with mean m and variance σ2 . Since the d.f. of X is
x−m
Φ ,
σ
Φ(qγs ) = γ.
qγ (X) ≈ m − 1.64σ.
EXAMPLE 3. There are n = 10 assets with random returns X1 , ..., Xn . The term “return”
means the income per $1 investment. For example, if the today price of a stock is $11,
11
while the yesterday price was $10, the return for this one-day period is 10 = 1.1. Note that
a return X may be less than one, and in this case we face a loss.
Assume X1 , ..., Xn to be independent and their distributions to be closely approximated
by the normal distribution with mean m and variance σ2 .
Let us compare two strategies of investing n million dollars: either investing the whole
sum in one asset, for example, in the first, or distributing the investment sum equally be-
tween n assets. We proceed from the VaR criterion with γ = 0.05.
For the first strategy, the income will be the r.v. Y1 = nX1 = 10X1 . The mean E{Y1 } = nm,
and Var{Y1 } = n2 σ2 , so to compute qγ (Y1 ) we should replace in (1.2.3) m by nm, and σ by
nσ. Replacing qγs by its approximate value −1.64, we have
For the second strategy, the income is the r.v. Y2 = X1 + ... + Xn . Hence, E{Y2 } = nm,
Var{Y2 } = nσ2 , and √
qγ (Y2 ) = mn − 1.64 nσ ≈ 10m − 5.2σ.
Thus, the second strategy is preferable, which might be expected from the very beginning.
Nevertheless, when the Xi ’s have a distribution different from normal, we may jump to a
different conclusion.
10 1. COMPARISON OF RANDOM VARIABLES
2 EXPECTED UTILITY
2.1 St. Petersburg’s paradox
The problem below was first investigated by Daniel Bernoulli in a paper published in
1738.
Consider a game of chance consisting of tossing a regular coin until a head appears.
Suppose that if the first head appears right away at the first toss, the payment equals 2 units
of money. If the first head appears at the second toss, the payment equals 4, and so on;
namely, if a head appears at the first time at the kth toss, the payment equals 2k .
It is easy to see that the payment in this case is a r.v. X taking values
2, 4, 8, ..., 2k , ... with probabilities
1 1 1 1
, , , ..., k , ... , respectively, and
2 4 8 2
1 1 1
E{X} = 2 · + 4 · + 8 · + ... = 1 + 1 + 1 + ... = ∞.
2 4 8
By the LLN, this means that if the game is played repeatedly, and X j is the payment in the
jth game, then with probability one
X1 + .. + Xn
→ ∞ as n → ∞.
n
Thus, in the long run, the average payment will be greater than an arbitrary large number.
Then if a player had proceeded from the LLN, she/he would have agreed to pay any,
arbitrary large, entry price for participating in each play.
2. Expected Utility 11
dx
du = k (2.2.1)
x
for a constant k. The solution to this equation is u(x) = k ln x + C, where C is another
constant. We will see soon that the values of k and C depend just on the choice of units in
measuring utility, and hence do not matter.
Consider now a random income X. In this case, the utility of the income is the r.v. u(X).
Bernoulli’s suggestion was to proceed from the expected utility E{u(X)}.
EXAMPLE 1. Assume that the utility function of the player in St. Petersburg’s paradox
is u(x) = ln x. Then the expected utility
∞ ∞ ∞
E{u(X)} = ∑ u(2k )2−k = ∑ ln(2k )2−k = (ln 2) ∑ k2−k = 2 ln 2,
k=1 k=1 k=1
X is equivalent to Y .
The investor who follows (2.2.2) is called an expected utility maximizer (EUM).
The first property of EUM criterion. The preference order (2.2.2) does not change
if u(x) is replaced by any function u∗ (x) = bu(x) + a, where b is a positive and a is an
arbitrary number.
EXAMPLE 1. Consider u(x) = k ln x + C as above. We see now that constants k and C
indeed do not matter, and we may restrict ourselves to u(x) = ln x.
2. Expected Utility 13
1
EXAMPLE 2. Consider u(x) = − , x ≥ 0 and u∗ (x) =
1 1+x
u (x)
*
x
u(x) + 1 = ; see Fig.2. The second function is positive but
0
x
1+x
u(x) reflects the same preference order as the first. The sign of u(x) does
-1 not matter; what matters when we compare X and Y is whether
E{u(X)} is larger than E{u(Y )} or not.
FIGURE 2. Now, assume that for a r.v. X, we can find a number c = c(X)
such that c ≃ X with respect to the order %.
The number c(X) so defined is called a certainty equivalent of X.
Now let us consider an EU maximizer with a utility function u. For such a person, the
above relation is equivalent to
E{u(X)} = u(c),
or
c(X) = u−1 (E{u(X)}),
1/α
∫b
1 1
c(X) = xα dx = b.
b (1 + α)1/α
0
Because (1 + α)1/α is decreasing in α, the smaller α, the smaller the certainty equiv-
alent. F
1
Since for α → 0, we have (1 + α)1/α → e, and in this case c(X) = b.
e
14 1. COMPARISON OF RANDOM VARIABLES
0 x
(a) (b)
FIGURE 3. Positive- and negative-power utility functions.
2. Negative-power functions: u(x) = −1/xα for all x > 0 and some α > 0; see Fig.3b.
We again deal only with positive r.v.’s, and E{u(X)} = −E{X −α }.
Both cases above may be described by the unified formula
1 1−γ
uγ (x) = x , γ ̸= 1. (2.2.3)
1−γ
In the case γ < 1, we have a positive power function, for γ > 1, we deal with a
negative power function.
3. The logarithmic utility function, u(x) = ln x, x > 0, is in a sense intermediate between
the two cases above and has been already discussed.
4. Quadratic utility functions: u(x) = 2ax − x2 , where parameter a > 0; the multiplier 2
is written for convenience. Certainly, we consider only r.v.’s X such that P(X ≤ a) =
1. Negative values of X are interpreted as the case when the investor loses or owes
money.
We have E{u(X)} = 2aE{X} − E{X 2 } = 2aE{X} + (E{X})2 −Var{X}. Thus, the
expected utility is a quadratic function of the mean and the variance.
5. Exponential utility functions. Let u(x) = −e−βx , where parameter β > 0, and the
function is considered for all x’s.
The graph is depicted in Fig.4.
The expected utility E{u(X)} = −E{e−βX } = −M(−β), where M(z) = E{ezX }.
The function M(z), we also use the notation MX (z) to emphasize that it depends on
the choice of the r.v. X, is the moment generating function of X.
In Exercises, we show that the certainty equivalent
1
c(X) = − ln(MX (−β)).
β
x
0
-1
u(x) = −e−βx
Proposition 1 Let u(x) = −e−βx and, under the EUM criterion with this utility function,
X % Y . Then w + X % w +Y for any number w.
The number w above may be interpreted as the initial wealth, and X and Y as random
incomes corresponding to two investment strategies. Proposition 1 claims that in the expo-
nential utility case, the preference relation between X and Y does not depend on the initial
wealth.
Proof of Proposition 1. By definition,
w + X % w +Y
iff
E{u(w + X)} ≥ E{u(w +Y )}.
For the particular u above,
1 1 2
2y − y2 ≥ 2 − = .
2 3 3
For this inequality to be true, we should have y ≥ 1 − √13 .
Hence, any acceptable premium G ≤ √1 ,
3
and Gmax = √1
3
≈ 0.57. .
2. Expected Utility 17
Next, we consider not an insured but an insurer. The latter offers the complete coverage
of a loss ξ for a premium H which, in general, may be different from G above. Assume that
the insurer is an EU maximizer with a utility function u1 (x) and a wealth of w1 . (Actually,
it is more natural to interpret w1 as an additional reserve kept by the insurer to fulfill its
obligations.) Following a similar logic, we obtain that an acceptable premium H for the
insurer must satisfy the inequality
(H + 1)α+1 − H α+1 = α + 1.
For example, when α = 1/2, it is easy to calculate—using even a simple calculator— that
Hmin ≈ 0.52.
For a premium P to be acceptable for both sides, the insurer and the insured, we should
have
Hmin ≤ P ≤ Gmax .
Hence, if Hmin > Gmax , insurance is impossible. If Hmin ≤ Gmax , the premium will be
chosen from the interval [Hmin , Gmax ]. For instance, in the situation of Examples 1-2, we
have 0.52 ≤ P ≤ 0.57.
18 1. COMPARISON OF RANDOM VARIABLES
Condition Z concerns an arbitrary preference order, not only the EUM criterion.
Proposition 2 Let % be a EUM order defined in (2.2.2). Then Condition Z holds iff
u(x) is concave.
Proposition 3 Let a r.v. X take on values from a finite or infinite interval I, and let a
function u(x) be concave on I. Then
Inequalities (2.4.1)-(2.4.2) are relevant to the basic question of insurance: why is it pos-
sible?
If the client is a risk averter (which is natural to assume since the client is willing to pay
to insure a risk), then u(x) is concave, and by Jensen’s inequality,
Gmax ≥ E{ξ}.
Thus, the maximum premium the client agrees to pay is larger than (or, for the boundary
case, equals) the average coverage of the risk, E{ξ}.
So, the company will get on the average more than it will pay, which means that the
company can function.
To the contrary, if the client had been a risk lover, from Jensen’s inequality it would have
followed that Gmax ≤ E{ξ}, and insurance would have been impossible.
20 1. COMPARISON OF RANDOM VARIABLES
E{r(X)} = λ. (3.1)
Once the premium g is completely specified by λ, the individual can choose any r(x)
provided that (3.1) is true.
EXAMPLE 1 (Proportional insurance). In this case,
r(x) = kx,
x
0
Consider a positive r.v. X, so its d.f. F(x) = 0 for x < 0. Let u(x) be a differentiable
function such that the integral ∫ ∞
u(x)dF(x)
0
exists and is finite. Set, also for simplicity,
u(0) = 0.
Then
∫ ∞
E{u(X)} = (1 − F(x))du(x). (3.2)
0
Setting u(x) = x, we obtain from (3.2) a useful formula for the expected value of a posi-
tive r.v.: ∫ ∞
E{X} = (1 − F(x))dx. (3.3)
0
In particular, from (3.3) it follows that the expected value equals the area between the
graph of F(x) and the line y = 1 .
22 1. COMPARISON OF RANDOM VARIABLES
where the number d is called a deductible. In this case, payment is carried out only if the
loss exceeds the level d, and if it happens, the insurer pays the overshoot.
Inserting (3.5) into (3.4), we have
∫ ∞
(1 − F0 (x))dx = λ. (3.6)
d
and m
d = ln m.
λ
Assume, for example that λ = 21 m. Then So, λ = m/2, so d = m ln 2 ≈ 0.69m.
EXAMPLE 3 (Insurance with a limit coverage). In this case,
x if x ≤ s,
r(x) =
s if x > s,
where s is the maximum the insurer will pay. Again using (3.4), we see that restriction (3.1)
may be written as ∫ s
(1 − F0 (x))dx = λ,
0
which is an equation for s.
We return to the optimization problem. Assume individual to be EUM, and her/his pref-
erences are preserved by ∫ ∞
U(F) = u(x)dF(x), (3.7)
0
and u is a non-decreasing utility function.
3. Optimal Payment from the Standpoint of an Insured 23
X(r) = w − g − X + r(X),
the wealth of the individual under the choice of a payment function r(x). Our goal is to
find a function r for which F(r) is the best. More rigorously, we a looking for a function r∗
which maximizes the function Q(r) = U(F(r) ).
24 1. COMPARISON OF RANDOM VARIABLES
Let us state it rigorously. For a fixed λ ∈ (0, m], consider the set of all function r(x)
satisfying (3.4), that is, the set
Theorem 4 Let u(x) in (3.7) be concave, and r∗ (x) = rd (x), where d satisfies (3.6).
Then for any function r(x) from R λ ,
So,
The optimal payment is the same for any concave utility function.
EXAMPLE 4. Two people facing the same loss X but having different utility functions, say
√
x and ln x, will prefer the same deductible policy and with the same deductible, provided
that they choose the same mean payment λ.