Chapter 3

Gambling, random walks and

the Central Limit Theorem

In this Chapter we will discuss the basic limit theorems in probability theory,
including the law of large numbers, central limit theorem and the theory of
large deviations. We will demonstrate the application of the law of large
numbers and CLT to the problem of Monte-Carlo calculations of integrals
and the theory of portfolio optimization. Using the simplest discrete portfolio
model we will explain the application of the theory of large deviations to risk
estimation. We will discuss the law of large numbers for Markov chains and
use it to analyse the simplest stochastic volatility model. We will then move
on to more complex stochastic volatility models to illustrate the limitations
of CLT in application to financial time series.
Finally we will analyse the discrete random walks in the context of the
Gambler’s ruin problem and apply the developed formalism to Cox-Ross-
Rubinstein model. In particular, we will derive Black-Scholes equation and
Black-Scholes formula for the value of a European call option in the contin-
uous limit of CRR model.

3.1 Random variables and laws of large num-

bers
To be able to understand the reasons for various limiting laws in probability
theory to hold, we will need some tools.

42
3.1.1 Useful probabilistic tools.
Chebyshev inequality gives a simple estimate for the probability of a devia-
tion from the mean:

Theorem 3.1.1. (Chebyshev Inequality) Let ξ be a random variable with

mean µ = E(ξ) and variance σ 2 = V ar(ξ). Then for any t > 0,

σ2
P r(|ξ − µ| ≥ t) ≤ .
t2
Proof:

σ 2 = V ar(ξ) ≡ E((ξ − µ)2 ) ≥ E(θ(|ξ − µ| − t)(ξ − µ)2 )

≥ t2 E(θ(|ξ − µ| − t)) = t2 P r(|ξ − µ| ≥ t).

A complementary form of Chebyshev inequality is

P r(|ξ − µ| < σt) ≥ 1 − t−2 .

Another useful inequality (due to Kolmogorov) estimates joint probabil-
ity for the sequence of sample sums to deviate from the respective expecta-
tions:

Theorem 3.1.2. Let ξ1 , . . . ξn be mutually independent random variables with

expectations µ1 , . . . , µn and variances σ12 , . . . , σn2 . Let

Sk = ξ 1 + . . . + ξ k

be the k-th sample sum. Let mk and sk be the expectation and the standard
deviation of Sk correspondingly.
For any t > 0,

P r(|S1 − m1 | < tsn , |S2 − m2 | < tsn , . . . , |Sn − mn | < tsn ) ≥ 1 − 1/t2

Kolmogorov inequality is stronger than Chebyshev inequality: the latter

would imply that the probability of the single event |Sn −mn | < tsn is at least
(1 − 1/tT2 ) while the former states that the probability of the intersection of
events nk=1 (|Sk − mk | < tsn ) satisfies the same bound.

43
3.1.2 Weak law of large numbers.
Theorem 3.1.3. Let {ξk } be the sequence of mutually independent identically
distributed variables. If the expectation µ = E(ξk ) exists, then for every ǫ > 0,

limn→∞ P r(|Sn /n − µ| > ǫ) = 0,

where Sn = nk=1 ξk is n-th sample sum. In other words, the probability that
P
sample average Sn /n differs from the expectation by less than an arbitrarily
prescribed ǫ tends to one.1
Proof: Note that the theorem does not require the existence of the vari-
ance of ξ. For the proof of the theorem in the complete generality, see Feller,
p. 231. Here we present the proof for the case σ 2 = V ar(ξ) < ∞: As
variances of independent variables add,

V ar(Sn ) = nσ 2 .

Therefore, V ar(Sn /n) = σ 2 /n. By Chebyshev inequality,

σ 2 n→∞
P r(|Sn /n − µ| > ǫ) ≤ −→ 0.
nǫ2

Note that weak convergence still allows for the set

(ξ1 (ω) − µ1 ) + · · · + (ξn (ω) − µn )
{ω| | ≥ δ}
n
to depend in an uncontrolled way on n.

3.1.3 Strong law of large numbers.

Let {ξi : Ω → R}i≥1 be the sequence of random variables with finite expec-
tations, µi = Eξi < ∞.
Def. The sequence of random variables ξ1 , ξ2 , . . . obeys the strong law of
large numbers if for any ǫ > 0, δ > 0, there is an N : for any r > 0 all r + 1
inequalities
(ξ1 − µ1 ) + · · · + (ξn − µn )
| | < ǫ, n = N, N + 1, . . . N + r, (3.1.1)
n
1
The weak law of large numbers deals with convergence in measure.

44
 (1 − δ) .2
will be satisfied with probability >P
We will use the notation Sn = nk=1 ξk for the sample sum, mn = E(Sn )
for the expectation value of the sample sum.
The following two theorems give sufficient conditions for a sequence to
satisfy the strong law of large numbers:

Theorem 3.1.4. (Kolmogorov Criterion.) If the series

X1
V ar(ξi )
i≥1
i2

converges, the sequence of mutually independent random variables {ξi }

obeys the strong law of large numbers.

Proof: Let Aν be the event that for at least one n: 2ν−1 < n ≤ 2ν , the
inequality (3.1.1) does not hold. We have to prove that for for ν > log2 (N )
and all r,
P r(Aν ) + P r(Aν+1 ) + . . . + P r(Aν+r ) < δ
P
In other words we need to prove the convergence of the series ν P r(Aν ).
By definition of Aν ,
|Sn − mn | ≥ nǫ > 2ν−1 ǫ,
for some n ∈ {2ν−1 + 1, . . . 2ν }. Hence by Kolmogorov inequality,

−2
P r(Aν ) ≤ 2ν−1 ǫ s22ν ,

where s22ν = V ar(S2ν ). Therefore,

∞ ∞ 2 ν ∞ ∞
X 4 X −2ν X 2 4 X 2 X −2ν 8 X σk2
P (Aν ) ≤ 2 2 σk = 2 σk 2 ≤ 2 2
< ∞,
ν=1
ǫ ν=1 k=1
ǫ k=1 ν
2 ≥k
ǫ k=1
k

where we estimated 2ν ≥k 2−2ν < 2/k 2 using an integral at the penultimate

P
step and applied condition of the theorem at the last.
The second theorem deals with identically distributed random variables:

Theorem 3.1.5. The sequence of i. i. d. random variables {Xk } such that

µ = E(Xk ) exists, obeys the strong law of large numbers.
2
In other words, the strong law asserts the convergence almost everywhere.

45
 Proof: Again, we would like to stress that the existence of variance is
not required for the validity of the Theorem. For the sake of brevity we will
present the proof under the extra assumption of the existence of variance,
V ar(ξk ) = σ 2 for all k. In this case the sequence
X σ2 X 1
2
2
= σ 2
<∞
k
k k
k
and the statement follows from Kolmogorov criterion.
It’s worth noting that Theorem 3.1.5 supersedes Theorem 3.1.3, as strong
convergence implies weak convergence.

3.2 Central Limit Theorem

The law of large numbers implies that if ξi : Ω → R are independent random
variables and with the same distribution, then
(ξ1 (ω) − µ1 ) + · · · + (ξn (ω) − µn )
lim =0
n→∞ n
for almost all ω ∈ Ω. The central limit theorem strengthens this by quanti-
fying the speed of convergence.
Theorem 3.2.1. Let ξi : Ω → R be i.i.d. random variables (independent
random variables with the same distribution) with µ = Eξi < ∞ and σ 2 =
V arξi < ∞ then
  
(ξ1 − µ) + · · · + (ξn − µ) σ σ
lim P ∈ [A √ , B √ ] =
n→∞ n n n
Z B
1 2
=√ e−x /2 dx.
2π A
Proof: We can’t prove CLT in full generality here. We can however
outline how it was discovered by Laplace and DeMoivre for the statistics of
coin tosses. Assume that we perform n tosses of an ’unfair’ coin such that
the probability of the head is p and the probability of the tail is q = 1 − p.
Let k be the number of heads observed. Let us introduce a random variable
ξ: ξ(Head) = 1, ξ(T ail) = 0. Then the random variable
n
X
Sn = ξk
k=1

46
counts the number of heads in our experiment.
We would like to estimate the probability
β  
X n
P rob(α ≤ Sn ≤ β) = pk q n−k (A)
k=α
k

in the limit n → ∞.
The law of large numbers implies that for large n the probability of ob-
serving the value of Sn far from pn is small. This motivates the introduction
of a new variable
k = np + δk
n


To estimate np+δ k
in the limit n → ∞, we use Stirling formula,
√
n! ∼ 2πnn+1/2 e−n , n >> 1.

The result is
  2
n 1 δk
pk q n−k ∼ √ e− 2nσ2 +Error ,
np + δk 2πnσ 2
δ3
where Error = C(p, q) nk2 + . . .. Here σ 2 = pq is the variance of ξ. For the
error to be small we must have
δk3
→ 0,
n2
as n → ∞. If this condition is satisfied for all terms in the expression (A), we
can conclude that the probability for the number of successes to fall between
α and β is given by the sum of Gaussian distribution functions!
Finally, let
δk
xk = √
σ n
√
Then, provided that x3k / n → 0 as n → ∞,
Sn − pn 1 √
P rob(xα ≤ √ ≤ xβ ) = √ (g(xα ) + g(xα + 1/ σ 2 n) + . . . + g(xβ )).
σ n σn
√ 2
Here g(x) = 1/ 2πe−x /2 . The above (Riemann) sum converges to the inte-
gral in the limit n → ∞ if we keep xα = A and xβ = B fixed, yielding the
statement of the theorem 3.2.1.

47
 The existence of the expectation, |E(ξ)| < ∞ is essential for the validity
of the weak law of large numbers. Consider for example Cauchy distribution
on R characterised by density
1 1
̺(x) = .
π 1 + x2
Let ξk be i. i. d. Cauchy random variable. It is possible to check that for
any n,
n
1 X Law
ξk ∼ ξ1 ,
n k=1
i. e. the empirical mean of any number of ξk ’s is still the Cauchy random
variable with the same distribution as ξ1 .
Let us illustrate CLT by computing the 95% confidence interval of the
sample mean
n
1X
Sn /n = ξi ,
n i=1
where {xi } is the sequence of i. i. d. random variables, E(ξi ) = µ, var(ξ) =
σ 2 . By definition 95% confidence interval is the interval symmetric around
the mean such the probability of random variable falling into this interval is
0.95.
If n is large, CLT applies and
  Z B
σ σ 1 2
P µ̄ − B √ ≤ Sn /n ≤ µ̄ + B √ ≈√ e−x /2 dx = 0.95
n n 2π −B

Resolving the last equality numerically, we get B ≈ 1.96. (Set Z ∼

N (0, 1) and note that P (|Z| ≤ x) = 0.95 for a number x which approximately
1.96). Therefore, the 95 percent confidence interval is
 
σ σ
µ̄ − 1.96 √ , µ̄ + 1.96 √
n n
More generally, if zα is the number such that P (|Z| ≤ z) = α for some
α > 0 then with 1 − α confidence one gets the above inequality with 1.96
replaced by z. (For example, α = 0.01% corresponds to zα ≈ 2.58.)

48
 It is remarkable that we can get a definite answer without the detailed
knowledge of the probability distribution of ξi ’s. Even if σ 2 is unknown we
can estimate its value using the law of large numbers
n
1 X
σ2 ≈ (ξi − µ̄)2 .
n − 1 i=1
If the ξi are i.i.d. then V ar(Sn ) = nσ 2 .
If ξi are independent random variables but do not have the same distri-
bution, then one needs some further conditions on the distributions of ξi in
order for the Central Limit Theorem to hold. One can weaken the assump-
tion about independence somewhat, but it is still important to have some
asymptotic independence.

3.3 The simplest applications of CLT and the

law of large numbers.
3.3.1 Monte Carlo Methods for integration
One way of integrating a function is to take averages of a randomly chosen
set of function values:
Let yi , i = 0, 1, . . . be i.i.d. with yi ∼ U (0, 1). Moreover, let f : [0, 1] → R
be a continuous function and write ξi = f (yi ). By the strong law of large
numbers for almost all ω,
n Z 1
1X
lim f (yi (ω)) = f (t) dt.
n→∞ n 0
i=1

The central limit theorem gives the following error bound, There is a 95%
probability that
(ξ1 + · · · + ξn )
n
lies between
Z 1 Z 1 
σ σ
f (t) dt − 1.96 √ , f (t) dt + 1.96 √
0 n 0 n
R1
where σ 2 is equal to 0 [f (t)]2 dt − ( f (t)dt)2 , and could again be estimated
R

from the sample . So in order to increase the precision (but again with the

49
same confidence) by a factor 10 one needs to increase the sample size by a
factor 100.
It is good to compare this with most simple classical numerical integra-
tion scheme (so not using a Monte Carlo approach). Let f : [0, 1] → R
be differentiable and such that |f ′ | ≤ M , and take yi be equally spaced:
yi = i/n, i = 0, 1, . . . , n. Then by the Mean-Value Theorem, f (x) ∈
[f (yi ) − M/n, f (yi ) + M/n] for each x ∈ [yi , yi+1 ]. So
Z yi+1
f (yi ) M M
f (t) dt ∈ + [− 2 , 2 ].
yi n n n

In fact, if f is twice differentiable then there exists a constant M so that

Z yi+1
f (yi ) + f (yi+1 ) M M
f (t) dt ∈ + [− 3 , 3 ].
yi 2n n n

Hence n
1
1X
Z
f (t) dt ∈ f (yi ) + [−M n−2 , Mn−2 ].
0 n i=1
This implies that using the non-probabilistic method, one increases the pre-
cision by a factor 100 by merely increasing the sample size by a factor 10. So
for such a nice function defined on an interval it does not really make sense
to use Monte Carlo methods for estimating the integral. In fact, any integra-
tion method used in practice approximates the integral by a trapezium rule,
quadratic interpolation, or even more advanced methods, and thus obtains
much better approximations than the above very naive method (provided the
function is sufficiently differentiable). But all these methods get less efficient
in higher dimensions: if f : [0, 1]d → R the error bounds get worse for large
d. For example, the analogous calculation, gives for the non-probabilistic
method an error bound of
Z 1 Z 1 n
1X
... f (t1 , . . . , td ) dt1 . . . dtd ∈ f (yi ) + [−M n−2/d , M n−2/d ].
0 0 n i=1

For large values of d these bounds become bad: the curse of dimensionality.
Compare this with the error of n−1/2 of the Monte-Carlo method. So if d is
large, the Monte-Carlo method wins.
So we see that Monte-Carlo methods for estimating integrals is some-
thing which is particularly useful when we work in higher dimensions, or

50
for functions which are not smooth. This is precisely the setting one has in
calculating stochastic integrals!
In the second half of the course various ways of improving on the simplest
Monte-Carlo method we discussed will be considered.

3.3.2 MATLAB exercises for Week 5.

1. Bernoulli experiment.

• Build the model of an unfair coin tossing experiment in MATLAB using ’binornd’
and ’cumsum’ commands.

• Construct a single plot which demonstrates the convergence of the sequence of

sample means to the expectation value for the various probabilities of a ’head’
p = 1/4, 1/2, 3/4. Is your plot consistent with the strong law of large numbers?
Try sample sizes N = 100, 1000, 10000.

• Investigate the fluctuations of the sample mean around the limiting value using
for sample size N = 1000: Use ’binopdf’ and ’hist’ commands to build the
probability distribution function of sample mean M1000 numerically and compare
your result against the prediction of central limit theorem for p = 1/4, 1/2, 3/4.

2. Monte-Carlo integration.

• Write your own MATLAB Monte-Carlo routine for calculating the integrals of
2
functions f (x) = x3 and f (x) = e−x over the interval [0, 1]. Analyse the conver-
gence of your routine as the function of the number of sampling points. Compare
your findings with the prediction of central limit theorem.

• Can you calculate the integral of f (x) = x−2/3 using Monte-Carlo method? Can
you use CLT to estimate the precision of your calculation? Explain your answers.

3.3.3 Analysing the value of the game.

Rules of the game: The gambler starts playing with some fixed sum of
M0 (pounds). At each step of the game, with equal probability, the gambler
either makes 80 percent or looses 60 percent. Assume that this game repeated
some large number n times. How well off do you expect the gambler to be
in the end of the game?

51
 Firstly, let us calculate the expected amount of money the gambler has
after n steps: Let ξk = 1 if the gambler wins in the k-th roundPand ξk = 0
n
if he loses. Then total number of wins in n rounds is Sn = k=1 ξk and
the total number of losses is n − Sn . Therefore, the amount of money the
gambler has after n rounds is
Mn = GSn Ln−Sn M0
pounds, where G = 1.8, L = 0.4. On average this is
!  "
S
n G n
E(Mn ) = M0 L E
L
As the rounds are independent, Mn is the product of independent random
variables,
n  ξk
Y G
Mn = L .
k=1
L
Therefore,
n
!  "
ξ  n
n
Y G k G+L
E(Mn ) = M0 L E = M0
k=1
L 2

Numerically,
E(Mn ) = M0 (1.1)n ,
meaning that the expected gain is growing exponentially. Imagine that a
casino offers this game for 1.09n M0 pounds for n-rounds, where M0 is the
amount the player wishes to start with. The argument for playing is that
the expected return will be quite big. For instance for n = 100 it is
(1.1n − 1.09n ) · M0 = 8.25 · 103 · M0 .
Should the gambler play this game?
Before answering ’you bet!’, consider the following argument: Let
 
1 Mn
rn = ln
n M0
be the average rate of increase of your bet. We can re-write this in terms of
the sample sum Sn as follows:
Sn
rn = ln(L) + ln(G/L) · (3.3.2)
n
52
The gambler has won the money if rn > 0 or
Sn ln(L)
>− ≈ 0.61.
n ln(G/L)

Now that is worrying: according to the strong law of large numbers, Snn
converges almost surely to 1/2 < 0.61! CLT allows us to quantify our worries:
the expected value of the sample mean Sn /n is 1/2. The standard deviation
is 2√1 n . CLT tells us that for large n, Sn /n will fall into the following interval
with probability 0.95:
√ √
[1/2 − 1.96/2/ n, 1/2 + 1.96/2/ n]

Using (3.3.2) we can re-state the above by saying that with probability 0.95,
√ √
−0.16 − 1.47/ n ≤ rn ≤ −0.16 + 1.47/ n.

In particular this means that with probability 0.95 the

√ gambler will be losing
money at the average rate of at least −0.16 + 1.47/ n < 0.
The situation seems paradoxical: given any ǫ > 0, we can find n large
enough so that the gambler will lose in the n-round game with probability
1 − ǫ. What about the large value of the expected sum in the end of the
game? It IS growing exponentially! The resolution is that with probability
less than ǫ a HUGE sum is won, making up for losses which happen most of
the time.
Is there any way to exploit the large expected value and win this game?
Surprisingly, we can use CLT to obtain the answer to this question as well:
Yes, there is if the gambler can diversif y and play this game independently
many times (either in parallel, or over a period of time - does not matter).
Really, assume that our gambler plays N 100-round games in parallel.
Let Mi be the amount of money the gambler receives from the i-th game.
Let
N
1 X
CN = Mi
N i=1
be the average amount of money the gambler receives per game. Note that

E(CN ) = ((G + L)/2)n ,

Due to the strong law of large numbers, CN is strongly converging to the

expected value ((G+L)/2)n , which means a very large profit for the gambler.

53
 We should however ask the question: how large should N be to ensure
that CN approaches ECN with probability close to one? The most crude
estimate following from CLT is that

V ar(CN ) = V ar(Mi )/N << E(CN )2

Equivalently,
n
E(Mi2 ) G2 + L 2

N >> −1= 2 − 1 ≈ 6 · 1014
(E(Mi ))2 (G + L)2

We are now ready to present our gambler with a mathematically sound

advice: yes, it is possible to exploit large mean values present in this game
and win a lot. You just need to play this game at least six hundred trillion
times. Oh and if you plan to start playing each game with a pound, do not
forget to bring a game fee of (1.09)100 · 6 · 1014 ≈ 3 · 1018 pounds with you -
it is due up front!
So our gambler is not ’he’ or ’she’ but ’it’ - Federal Reserve or the People’s
Bank of China - managing risk to its investment via diversification.
Of course, the numbers used in the above example are greatly exagger-
ated to demonstrate an important point: the expectation values can easily
mislead. Note however how easy and effective our analysis was, given the
probabilistic tools we mastered - the law of large numbers and central limit
theorem.

3.3.4 Portfolio optimization via diversification.

As we have learned in the previous section, diversification can be used to
manage even the wildest fluctuations of luck, given sufficient resources.
This consequence of the law of large numbers is routinely used to build
investment portfolios which maximize the probability of a return falling into
a desired range. The main idea is to diversif y the portfolio by including
many independent assets. There are two main types of diversification: ver-
tical and horizontal.

Background remark.(Source: Wikipedia)

Horizontal Diversification.

54
Horizontal diversification is when you diversify between same-type invest-
ments. It can be a broad diversification (like investing in several NASDAQ
companies) or more narrowed (investing in several stocks of the same branch
or sector).

Vertical Diversification.

Vertical Diversification is investing between different types of investment.

Again, it can be a very broad diversification, like diversifying between bonds
and stocks, or a more narrowed diversification, like diversifying between stocks
of different branches.

While horizontal diversification lessens the risk of just investing all-in-one,

a vertical diversification goes far beyond that and insures you against market
and/or economical changes. Furthermore, the broader the diversification the
lesser the risk.
Let us discuss the mathematics of diversification. Assume that our port-
folio contains n types of assets (shares, bonds, etc.) each characterised by a
return r1 , r2 , . . . , rn (For instance, ri can denote monthly increase in the price
of a share). Note that ri is a random variable. If these random variables are
independent, the portfolio is called diversified. Let

µi = E(ri ), σi2 = V ar(ri )

be the expectation value and the variance of ri correspondingly. To avoid

unnecessary technicalities, we will assume that all expectations are equal,

µk = µ > 0 ∀k.
As we are interested in the properties of our portfolio for large values of
n, we need some extra assumptions to control the set of our random variables
in the limit n → ∞. We will use the simplest assumption - that all returns
are uniformly bounded by an n-independent constant,

|ri | < V ∀i ∈ N. (3.3.3)

We also need to exclude cases σk = 0 from consideration: if there is such

a thing as an asset with zero volatility (standard deviation) and a positive
return, everybody should just build their portfolios out of this asset only and

55
that’s it. (Which proves by the way, that such an asset does not exist.) We
will assume that all sigma’s are uniformly bounded away from zero,
0 < v < σk2 , ∀k. (3.3.4)
Let Ni be the numbers of assets of type i in the portfolio, 1 ≤ i ≤ n. Let
n
X
N= Nk .
k=1

Then the average return from our portfolio is

n
1 X
Rn = Nk rk
N k=1

Note that in the terminology of probability theory, the average return is just
the empirical mean of all returns. The expected value of the average return
is n
1 X
E(Rn ) = Nk µ = µ,
N k=1
the variance is n
1 X 2 2
V ar(Rn ) = 2 N σ .
N k=1 k k
The simplest way to demonstrate the advantage of portfolio diversification
is to consider the case
N1 = N2 = . . . Nn = K,
where K does not depend on n. In other words we assume that our portfolio
contains the equal number K of assets of each type. In this case
n
1 X 2 1 n→∞
V ar(Rn ) = 2 σk < (2V )2 −→ 0.
n k=1 n

Therefore, we can hope that in the large-n limit, we will be getting a

guaranteed return µ from our portfolio. Indeed, the sequence of independent
random variables Kr1 , Kr2 , . . . , Krn satisfies Kolmogorov criterion:
n n
X V arKrn X 4K 2 V 2
≤ < ∞.
k=1
k2 k=1
k2

56
Therefore, Rn strongly converges to µ as n → ∞.
As n = ∞ is unrealistic, we also need to know the statistics of fluctuations
of Rn around the mean value. This information is provided by central limit
theorem. Here is the slight modification of CLT for independent, but not
necessarily identically distributed variables:
Theorem 3.3.1. (Lindeberg) Every uniformly bounded (i. e. satisfying
(3.3.3)) sequence X1 , X2 , . . . of mutually independent random variables obeys
Z β
Sn − µn dx − x2
P r(α < < β) = e 2,
σn α 2π

provided that σn → ∞ as n → ∞. Here Sn = nk=1 Xk is the sample sum,

P
µn = E(Sn ), σn2 = V ar(Sn ).
Note that the condition σn → ∞ as n → ∞ ensures that cases like
X2 = X3 = . . . 0 for which the above statement is clearly wrong are excluded.
CLT allows us to answer many important questions about our portfolio
such as: what is the 95 percent confidence interval for Rn ? Does it contain
zero?
Now, let us relax the assumption Ni = Nj which will allow us to consider
the problem of portfolio optimization. The typical aim of portfolio optimiza-
tion is to minimize the variance around some pre-determined expected return
by choosing N1 , N2 , . . . Nn appropriately. The corresponding mathematical
formulation is called Markowitz problem3 :

Find Ni ’s minimizing V ar(Rn ) given the constraints

n
X
Nk = N, (3.3.5)
k=1
E(Rn ) = M. (3.3.6)

In our case, the second constraint is absent as E(Rn ) = µ for any choice
of Nk ’s. The minimization problem of V ar(Rn ) subject to the constraint
P
k Nk = N is an example of the classical quadratic optimization problem.
3
See Yuh-Dauh Lyuu, Financial Engineering and Computation, for more details. Harry
Markowitz developed the modern portfolio theory in 1950’s. In 1990 he was awarded Nobel
prize in economics.

57
For the case at hand, this problem has a unique solution (as V ar(Rn ) is a
convex function of Nk ’s and there is at least one set of positive Nk ’s solving
the constraint) which can be found using the method of Lagrange multipliers.
Let X
F (N1 , . . . Nk , λ) = V ar(Rn ) − λ( Nk − N ).
k

The critical point of F satisfies

2Nk
σk2 = λ, k = 1, 2, . . . , n.
N2
Therefore, the optimal distribution of our portfolio over assets is given by

(0) λN 2
Nk = .
2σk2
Note that the optimal portfolio tends to have more assets with smaller vari-
ance. Lagrange multiplier λ can be computed from the constraint (3.3.5)
giving us
(0) N
Nk = Pn σ −2
−2 k
σ
m=1 m
The minimal variance is
n
1 X (0) 2 2 1
min(V ar(Rn )) = Nk σ k = P n .
N 2 k=1 −2
m=1 σm

Using (3.3.3), we can immediately verify that the optimized variance van-
ishes in the limit of large portfolio size:
(2V )2
min(V ar(Rn )) ≤ → 0, for n → ∞.
n
Fantastic, we managed to optimize our portfolio and the resulting variance
of the mean return becomes small as the number of assets grows. Moreover,
in virtue of the theorem 3.3.1, the fluctuations
√ of Rn around the mean value
µ in the interval of length of order 1/ n are Gaussian with variance equal
to min(V ar(Rn )) computed above.4
4
My friend from Moscow ’City’ said this about Markowitz portfolio theory: ’As a quan-
titative theory it never works for realistic portfolio sizes as something is always happening
which invalidates CLT. But thanks God for Markowitz as now we can explain to customers
why their portfolios have so many under-performing assets.’

58
 It is important to remember that both the law of large numbers and
the central limit theorem require mutual independence of random variables.
Portfolio theory provides an example of the spectacular violation of these
limit theorems due to correlations between returns from the various assets.
Let
σij = E ((ri − E(ri ))(rj − E(rj ))) , i 6= j
be the covariance of the returns of assets i and j. Then
! n "
1 X 2 2 X
V ar(Rn ) = 2 Nk σk + Ni Nj σij
N k=1 i6=j

The second term on the right hand side of the above expression is called
(somewhat misleadingly) the systematic portfolio risk. The first term is
referred to as the special or non-systematic risk. The systematic risk is
determined by the correlations between distinct assets in the portfolio.
To make the point, let us assume that Nk = K for all k. Let us also
assume that all variances are equal to each other and to σ 2 . Finally, let us
assume that all assets are correlated in such a way that

σij = zσ 2 ,

where z ≤ 1 is the correlation coefficient. Then

σ 2 n(n − 1) 2
V ar(Rn ) = + zσ .
n n2
Note that non-systematic risk vanishes in the limit of large n, whereas sys-
tematic risk converges to the limit equal to the mean covariance for all pairs
of assets (Markowitz law of mean covariation). It is clear that in the
presence of mean covariance, the sequence Rn does not converge to µ. As a
result, it is impossible to achieve perfect diversification of the corresponding
portfolio.
It is still possible to prove limiting theorems for non-independent random
variables given that the correlations are weak in some sense. A typical ex-
ample is served by the infinite sequence of correlated random variables which
can be broken into subsets of fixed size such that elements belonging to dif-
ferent subsets are mutually independent. We are not going to pursue the
limiting theorems for correlated variables in this course.

59
3.4 Risk estimation and the theory of large
deviations.
Despite the demonstrated usefulness of the law of large numbers and central
limit theorems for portfolio optimization and game analysis there are some
important questions which these theorems do not address. Assume for ex-
ample that we have a well diversified portfolio √ consisting of n assets with
expected return µ and standard deviation C/ n. Central limit theorem al-
lows us
√ to estimate the probability of the return to be in the interval of size
∼ 1/ n around the mean value. However, it cannot be used to estimate
the probability of a large loss L, Rn < −L in the limit of large n. Which
is unfortunate, as this probability is of principle importance for contingency
planning for rare but catastrophic events.
The probability of large losses can often be effectively estimated using the
theory of large deviations. To illustrate this theory we will use the following
simple example:
Assume that out portfolio consists on n independent assets. Each asset
brings us one pound per month with probability p > 1/2 and loses us one
pound with probability q = 1 − p < 1/2. Therefore, we can earn up to n
pounds and loose up to n pounds per month. The expected mean return is
PIf ri is the return from the i-th asset, the mean return per
however positive:
asset is Rn = n1 k rk and

E(Rn ) = (p − q) > 0.

Due to the law of large numbers, Rn stays close to the deterministic value
p − q with probability close to 1.
However, what is the probability of a large loss L = τ n, where 0 ≤ τ < 1
is fixed?
The answer is given by the following result:
Proposition 3.4.1. (Cramer) Let Pr1 , r2 , . . . , rn be i. i. d. bounded random
variables, µ = E(rk ). Let Rn = n1 nk=1 rk . Fix τ > −µ. Then
1
lim ln(P rob(−Rn > τ )) = −I(τ ),
n→∞ n

where I(τ ) is called the rate function or Cramer function,

I(τ ) = supθ>0 (θτ − ln(E(e−θr ))).

60
In other words, I(τ ) is the Legendre transform of the cumulant generating
function lnE(e−θr ).

The appearance of the rate function in the above form is easy to follow
using Chernof f inequality. By definition,

P rob(−Rn > τ ) = E(χ(−Rn − τ )),

where χ(x) is 1 for x ≥ 0 and is zero otherwise. Therefore, for any positive
θ,
P rob(−Rn > τ ) ≤ E(enθ(−Rn −τ ) χ(−Rn − τ )) ≤ E(enθ(−Rn −τ ) )
n
−θr ))
Y
= e−nθτ E(e−θrk ) = e−n(θτ −lnE(e .
k=1

To find the tightest bound, we need to find the minimum of the right hand
side over all θ > 0. Therefore,
−θr ))
P rob(Rn < −τ ) ≤ e−nsupθ>0 (θτ −lnE(e = e−nI(τ ) .

Taking the logarithm of both sides,

lnP rob(Rn < −τ )
≤ −I(τ ).
n
The derived inequality of often referred to as Chernoff-Cramer bound.
Cramer theorem gives sufficient conditions for Chernoff-Cramer bound to
be tight. The proof is based on the technique of asymptotic expansion of
certain integrals due to Laplace which cannot be discussed here.
For large n, the above proposition states that, with logarithmic precision,

P rob(Rn < −τ ) ≈ e−N I(τ )

Therefore, to estimate the risk of our ’discrete’ portfolio we need to evaluate

the rate function. Firstly,

E(e−θr ) = pe−θ + qeθ .

The function θτ − ln(E(e−θr )) is maximized at the point θc :

−pe−θc + qeθc
τ= .
pe−θc + qeθc

61
This equation can be solved leading to a remarkable answer for the rate
function:   
1−τ 1+τ
I(τ ) = DKL , ||(p, q) ,
2 2
where DKL is Kullback-Leibler divergence or relative entropy. DKL is widely
used in information theory and statistical mechanics as a measure of differ-
ence between two stochastic vectors. If P~ and Q ~ are d-dimensional stochastic
vectors,
d  
~ ~
X Pk
DKL (P ||Q) = Pk ln .
k=1
Q k

For d = 2 the explicit expression is

      
1−τ 1+τ 1−τ 1−τ 1+τ 1+τ
DKL , ||(p, q) = ln + ln .
2 2 2 2p 2 2q
Note that DKL is not symmetric,
DKL (P~ ||Q)
~ 6= DKL (Q||
~ P~ ).

Relative entropy possesses the following remarkable property (Gibbs in-

equality): for any P~ and Q,
~

DKL (P~ ||Q)

~ ≥ 0.

Moreover, DKL (P~ ||Q)

~ = 0 iff P~ = Q.
~
You will have an opportunity to test the large deviations result we derived
during the next week’s MATLAB session.

Remarks. Harald Cramer was a Swedish mathematician. He originated

the theory of large deviations while working as a consultant for an insurance
company. The problem of risk management was therefore the original in-
spiration for this theory! Note that Cramer’s theorem consists of two parts:
firstly, it states that P rob(Mn > x) decays exponentially with n. Secondly,
it states that the rate of this decay, the rate function, can be calculated in
terms of the cumulant generating function. It often the case that the proba-
bility distribution of random variables contributing to Mn is not known. In
this case the rate function cannot be computed theoretically but the state-
ment concerning the asymptotically linear dependence of ln(P rob(Mn > x))
on n may still hold true and provide useful information about the probability
of rare events.

62
3.4.1 Week 6 MATLAB exercises.
Numerical study of the Game.
Use the binomial distribution to analyse the n = 100-round game introduced
in Sec. 3.3.3:    i  n−i
n 1 1
P rob.(Sn = i) =
i 2 2

• Write a MATLAB programme for the computation the chance of loss, expected
gain and gain variance in Matlab. Compare your answer with theoretical pre-
dictions. How does the standard deviation compare with the expected value?
Remark. If you can use the statistics package, then you can use the operator
binopdf - which enables to choose larger values of n.

• Write a MATLAB programme which computes the expected value of the mean
rate of winning, its variance and build its histogram. Perform a careful compar-
ison of your numerical results with the predictions of the law of large numbers
and central limit theorem.

Risk estimation for the discrete portfolio model.

• Write MATLAB programme for the calculation of the probability of loss (τ = 0)
for the discrete portfolio model discussed in Sec. 3.4 based on the binomial
distribution. You can use p = 0.6 in your calculations. Plot this probability as
a function of n using the semi-log scale. Plot the theoretical prediction based on
the large deviations theory on the same plot. How do the two curves compare?

• Calculate the same probability using the (incorrectly applied) central limit the-
orem. How does the answer compare with the exact result?

3.4.2 An example of numerical investigation of CLT

and the law of large numbers for independent
Bernoulli trials.
CLT
%1. ’Verify’ LLN and CLT for the sequence of tosses of a fair coin.
%1.1 Convergence of sample means
%Fix the number of trials

63
n=2000;
%Create sample sequence
x=round(rand(1,n));
%Create the sequence of sample means
sm=zeros(1,n); for k=1:n
sm(k)=sum(x(1:k))/k;
end
%Plot the sequence and the expected value it should converge to
xx=1:n; subplot(2,2,1), plot(xx,1/2*ones(size(xx,2)),’k--’,
xx,sm,’r-’)
legend(’The limit’,’Random sequence of sample sums’)
%Conclusion: the sequence of sample sums seems to converge to 1/2.
yes=input(’Continue?’)
%1.2. Quantitative study of convergence
%Fix the number of experiments
expers=2000;
%Create the sequence of sample sums S_n’s
Sfinal=zeros(1,expers);
for k=1:expers
Sfinal(k)=sum(round(rand(1,n)))/n;
end

%Calculate experimental variance

var_exp=var(Sfinal)
%Calculate theoretical variance
var_th=1/2*(1-1/2)/n
%How close var_th and var_exp are?
Error=abs(var_exp-var_th)/var_exp

%Build the histogram of S_n’s

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%Construct the scaled variable

Sscaled=(Sfinal-1/2)/sqrt(var_th); step=0.4; dev=[-2:step:2];

pdf_exp=hist(Sscaled,dev); pdf_exp=pdf_exp/sum(pdf_exp);
pdf_th=exp(-dev.^2/2)/sqrt(2*pi); pdf_th=pdf_th/sum(pdf_th);
subplot(2,2,4), plot(dev,pdf_th,’r--’, dev, pdf_exp,’k-’)
legend(’CLT’,’Experiment’)

64
The probability of loss and large deviations theorem.
%Theory
tau=0; p=0.8; ptau=[(1-tau)/2,(1+tau)/2]; pp=[p,1-p];
dkl=sum(ptau.*log(ptau./pp))
%Max number of rounds
N=3000; n=1:N; prob=exp(-dkl*n);

%Experiment
%number of -1’s is greater than the number of ones.
for k=1:N prob_exp(k)=sum(binopdf(0:ceil(k/2-1),k,p)); end

%CLT
prob_clt=1/2*erfc((2*p-1)/(8*p*(1-p))^(1/2).*(n.^(1/2)));

subplot(2,2,1),
plot(10:100,log(prob(10:100))./(10:100),’r--’,10:100,
log(prob_exp(10:100))./(10:100),’k-’,10:100,
log(prob_clt(10:100))./(10:100),’b-.’);
legend(’Large deviations’,’Experiment’,’CLT’),xlabel(’Number of
rounds, n’),ylabel(’log(Pr(Loss))/n’) title(’log(Pr(Loss))/n for a
small number of rounds, p=0.8’) LA=[N/2:N];

subplot(2,2,2),
plot(LA,log(prob(LA))./LA,’r--’,LA,log(prob_exp(LA))./LA,’k-’,
LA,log(prob_clt(LA))./LA,’b-.’);
legend(’Large deviations’,’Experiment’,’CLT’),xlabel(’Number of
rounds, n’),ylabel(’log(Pr(Loss))/n’) title(’log(Pr(Loss))/n for a
large number of rounds, p=0.8’)

subplot(2,2,3),
semilogy(10:100,prob(10:100),’r--’,10:100,prob_exp(10:100),’k-’,
10:100,prob_clt(10:100),’b-.’);
legend(’Large deviations’,’Experiment’,’CLT’),xlabel(’Number of

65
rounds, n’),ylabel(’Pr(Loss)’) title(’Pr(Loss) for a small number of
rounds, p=0.8’) LA=[N/2:N];

subplot(2,2,4),
semilogy(LA,prob(LA),’r--’,LA,prob_exp(LA),’k-’,LA,prob_clt(LA),’b-.’);
legend(’Large deviations’,’Experiment’,’CLT’),xlabel(’Number of
rounds, n’),ylabel(’Pr(Loss)’) title(’Pr(Loss) for a large number of
rounds, p=0.8’)

log(Pr(Loss))/n for a small number of rounds, p=0.8 log(Pr(Loss))/n for a large number of rounds, p=0.8
−0.2 −0.22
Large deviations Large deviations
−0.25 Experiment −0.23 Experiment
CLT CLT
−0.3 −0.24
log(Pr(Loss))/n

log(Pr(Loss))/n

−0.35 −0.25

−0.4 −0.26

−0.45 −0.27

−0.5 −0.28

−0.55 −0.29
0 20 40 60 80 100 1500 2000 2500 3000
Number of rounds, n Number of rounds, n

Pr(Loss) for a small number of rounds, p=0.8 Pr(Loss) for a large number of rounds, p=0.8
0 −100
10 10
Large deviations Large deviations
Experiment Experiment
CLT CLT
−5
10
Pr(Loss)

Pr(Loss)

−200
10

−10
10

−300
10
−15
10
0 20 40 60 80 100 1500 2000 2500 3000
Number of rounds, n

Figure 3.1: The probability of loss: experiment, CLT and large deviations.

66
3.5 The law of large numbers for Markov chains
The assumption of independence of logarithmic price increments of past price
movements plays a fundamental role in financial mathematics.5 According
to such an assumption, the logarithm of a share’s price on day n conditional
on the logarithm of the price on day n − 1 is independent of the rest of the
price history (i. e. prices on days n − 2, n − 3, . . .). Markov chains are
mathematical models describing such random processes.
Let ξt be sequence of discrete random variables taking values in {1, . . . , n},
where t ∈ N is ’discrete time’. If

P r(ξt+1 = j|ξt = i, ξt−1 = k1 , ξt−2 = k2 , . . .) = P r(ξt+1 = j|ξt = i),

the sequence ξt is called a Markov chain. Matrix

pji = P r(ξt+1 = j|ξt = i)

is often referred to as transition matrix. If the random process ξi is stationary,

the Markovian property is equivalent to time-independence of the transition
matrix.
The transition matrix P = (pji )1≤i,j≤n is a stochastic matrix. Therefore,
we can use the results of Chapter 1 to study the statistics of Markov chains.
If in particular, P is irreducible (i.e. there is k: P k has only strictly positive
entries) then
lim P n → (~q ~q . . . ~q)
n→∞

where ~q is the unique column eigenvector of P associated with eigenvalue

1 (Perron-Frobenius theorem for stochastic matrices). Recall that ~q is also
referred to as the equilibrium state. Let qi be the i-th component of vector
~q.
Due to the simplicity of the limiting form of P n , the equilibrium statistics
of irreducible Markov chains are easy to study. For example, let Tni be the
number of times that the state i ∈ {1, . . . , n} has occurred in a sequence
ξ1 , . . . , ξk and Tnij be the number of times that the transition ij appears in
the sequence ξ1 , . . . , ξk .
5
This assumption can be thought of as a consequence of the ’efficient market hypoth-
esis’, according to which the present price of a security encompasses all the available
information concerning this security.

67
Theorem 3.5.1. (Law of Large numbers for Markov chains)
 i 
Tn
P | − qi | ≥ ǫ → 0, n → ∞,
n
 ij 
Tn
P | − Pji qi | ≥ ǫ → 0, n → ∞.
n
i
We will not prove this theorem here. Note only, that Tnn is the frequency
with which the value i appears in the sample sequence, whereas qi is the
probability of such an event in the equilibrium state. Therefore, we have
a typical weak law of large numbers: frequency converges to probability.
Similarly, the probability of the transition i → j is P r(ξn = j, ξn−1 = i) =
ij
P r(ξn = j | ξn−1 = i)P r(ξn−1 = i) = Pji qi , whereas Tnn is its frequency.
Notice also that the rate of convergence of frequency to probability stated
in the above theorem depends on the size of the second largest eigenvalue of
P.

3.5.1 The Markov model for crude oil data.

In this section we will apply Markov models to analyse the price fluctuations
of crude oil traded on the New York Mercantile Exchange in 1995-1997.6 .
Let P (n) be the oil price on the n-th day. Let

L(n) = ln(P (n)).

Let
D(n) = L(n) − L(n − 1)
be the logarithmic daily price increment. Let us define the following four
states of the oil market:

State Range of increment

1 D(n) ≤ −0.01
2 −0.01 ≤ D(n) ≤ 0
3 0 ≤ D(n) ≤ 0.01
4 D(n) ≥ 0.01
6
See Sheldon M. Ross, An elementary introduction to mathematical finance, Chapter
12

68
 In other words, state 1 describes a large (greater than one percent drop)
in crude oil prices, state 2 corresponds to the price drop between zero and
one percent, state 3 describes the price rise between zero and one percent.
Finally, state 4 describes a large (over one percent) increase in the oil price.
The following table shows the observed frequency of the state S(n) given
the state S(n − 1):

S(n) \ S(n − 1) 1 2 3 4
1 0.31 0.21 0.15 0.27
2 0.23 0.30 0.28 0.32
3 0.25 0.21 0.28 0.16
4 0.21 0.28 0.29 0.25
Assuming, that D(n) is independent on D(n−2), D(n−3), . . ., given D(n−1),
we arrive at the description of the fluctuation of oil prices in terms of a four-
state Markov chain with the following transition matrix:
 
0.31 0.21 0.15 0.27
 0.23 0.30 0.28 0.32 
P =  0.25 0.21 0.28 0.16 


0.21 0.28 0.29 0.25

Recall that the elements of the i-th column of P are equal to probabilities
of S(n) = j given S(n − 1) = i. Note that all columns are different and the
difference is significant in the statistical sense. This means that the statistics
of D(n) are not independent of D(n − 1). In other words, oil prices do
not follow geometric Brownian motion and the corresponding Black-Scholes
theory which we will discuss in the following Chapters does not apply. The
model above is an example of the class of stochastic volatility models which
allow the price increment variance to depend on the current state of the
market.
The analysis of the Markov model at hand is pretty straightforward: the
stochastic matrix P defined above is irreducible. Therefore, there is a unique
equilibrium state, which we can find using MATLAB:
 
0.2357
~ =  0.2842 
 
Q  0.2221 
0.2580

69
The norm of the second largest eigenvalue is 0.1 signalling a rapid convergence
to equilibrium. According to Theorem 3.5.1, matrix elements of Q ~ give us the
frequency of large and medium price drops and medium and large price rises
correspondingly. Note that the most frequent state is that of the moderate
price drop closely followed by the state of the large price increase. The matrix
of joint probabilities P r(S(n) = j, S(n − 1) = i) in the equilibrium state is
 
0.0731 0.0597 0.0333 0.0697
~ =  0.0542 0.0853 0.0622 0.0826 
 
P ⊗Q  0.0589 0.0597 0.0622 0.0413 
0.0495 0.0796 0.0644 0.0645

Note that this matrix is non-symmetric which reflects the dependence of

random variables S(n) and S(n − 1). As we can see the most frequent
transition (frequency 0.0853) is 2 → 2 - from the state of the moderate price
drop to itself. The least frequent transition (frequency 0.0333) is from the
state of moderate price increase to the state of the large price drop. Note
that the frequency of the transition from the large price rise to the large
price drop is twice as large, reflecting large fluctuations in the size of price
increments for this market.
It would be quite interesting to perform a similar analysis on the modern
day crude oil prices data.

3.6 FTSE 100: clustering, long range corre-

lations, GARCH.
According to CLT, sample means of the large number n of independent ran-
dom variables are described by a Gaussian distribution
√ with the mean equal
to the expectation value and variance of order 1/ n. For the CLT to hold,
the random variables contributing to the sum should be well behaved, for
instance their variances should be finite.
The ideology of CLT is so popular, that people tend to model any ran-
dom variable (e.g. index price) which depends on many random parameters
using Gaussian distributions. Unfortunately, it turns out that some very
successful empirical models used to explain the observed properties of price
movements- stochastic nature of volatility (standard deviation of logarith-
mic price increments), clustering and long range temporal correlations lead

70
to power law distributions of price increments. Such a behavior clearly vio-
lates CLT. The possible reason for such a violation is the presence of strong
correlations between stock prices contributing to a given index.
To motivate such models, let us consider the volatility of FTSE100 index
shown in Fig. 3.6 7 :

Figure 3.2: Bloomberg data for daily volatility fluctuations.

Notice that periods of high volatility tend to cluster together. Notice also
that there are long periods of volatility staying below the mean value -the
periods of mean reversion.
Given such a behaviour, it is quite improbable that daily volatilities are
described by independent random variables. Indeed, let

hσ 2 (t)σ 2 (t + τ )i
A(t, τ ) = 2 −1
hσ (t)ihσ 2 (t + τ )i

be the normalized covariance between variances on day t and day t + τ .

Parameter τ measures time lag between the measurements. Function A(t, τ )
7
The graphs of the original data used in this section are taken from
Alessio Farhadi, ARCH and GARCH models for forecasting volatility, http :
//www.quantnotes.com/f undamentals/basics/archgarch.htm.

71
is often referred to as autocorrelation function. For a stationary sequence
of random variables, the autocorrelation function depends on the time lag τ
only.
If volatilities on different days were independent random variables, the
autocorrelation function would have been zero. Look however at Fig. 3.6:
not only A(τ ) 6= 0, it also decays very slowly as the time lag increases: the
red curve shows the result of LMS fit of the exponential function e−τ /τ 0 to
data. The law of decay is much better described by a power law, signalling
long memory in the sequence of daily volatilities.

Figure 3.3: Volatility-volatility covariance as the function of time lag.

The observed behaviour of volatility (clustering, mean reversion, long

memory) enables volatility forecasting: the most likely volatility tomorrow,
will depend on the chain of observed volatilities up to date. This makes
volatility very different from price fluctuations which cannot be forecasted.8
To make the volatility prediction feasible, people looked for simple proba-
bilistic models of stochastic volatility which will have the observed properties
of the volatility time series. The simplest model has been discussed in the
8
This property of volatility makes volatility trading an exciting field for applications
of financial mathematics, see http://en.wikipedia.org/wiki/VIX for more information on
S&P 500 volatility index VIX.

72
previous Section: transition probabilities from state i to state j depended on
the state i. For this Markov chain we did not attempt to model the stochastic
volatility explicitly. Instead, transition probabilities were estimated directly
from data.
The first explicit stochastic volatility model is ARCH(m) model invented
in 1982 by Engle to model the volatility of UK inflation. Let D(n) be the
logarithmic price increment on day n, let σn be the volatility of D(n). The
variance is equal to σn2 . According to Engle,

D(n) = σ(n)W (n), (3.6.7)

m
X
2
σ (n) = γV0 + αk D2 (n − k). (3.6.8)
k=1

Here V0 is the expected variance, {W (n) ∼ N (0, 1)} are independent random
variables; parameters γ and α’s are non-negative numbers such that
m
X
γ+ αk = 1.
k=1

ARCH stands for ’autoregressive conditional heteroskedasticity’:

(i) Heteroskedasticity refers to the fact that volatility is itself volatile. In

this sense any stochastic volatility model can be called heteroskedastic.9
(ii) The model is autoregressive as variance on day n is a linear combination
(regression) of squares of past price
Pincrements.
Note that the constraint γ + m k=1 αk = 1 follows from the definition
of V0 as the expected variance: if we average (3.6.8) and use the fact that
E(D(n)2 ) = E(σ(n)2 ) = V0 , we will get
m
X
V0 = (γ + αk )V0 ,
k=1

which leads to the above constraint. Note than for any ARCH model,
E(D(n)D(n − k)) = 0 for k 6= 0. Therefore price increments are uncorrelated
(but not independent - dependence creeps in via dependencies in volatilities).
It is easy to see how ARCH can lead to clamping and mean reversion: large
9
’hetero’ literally translates as ’different’, ’skeda’ - as ’step’ from Greek.

73
values of observed price increments lead to large future volatility, which in
turn increases the probability of large future price increments. If on the other
hand, a sequence of small price increments has occurred, the predicted future
volatility will be less than the mean value.
The most common model in the family is ARCH(1):
D(n) = σ(n)W (n), (3.6.9)
σ (n) = γV0 + (1 − γ)D2 (n − 1),
2
(3.6.10)
which can be easily studied using MATLAB. Here is the script generating
the series of volatilities according to ARCH(1):
%%%%%%%%%%%%%%%%%%%%%%
%ARCH(1)
%%%%%%%%%%%%%%%%%%%%%%
%Expected variance
V0=1;
%Runlength
N=1000000;
%Variance array
V=V0*ones(1,N);
%Increments array
D=zeros(1,N);
%Memory parameter
gamma=0.4;

for k=2:N
V(k)=gamma*V0+(1-gamma)*D(k-1)^2;
D(k)=sqrt(V(k))*randn;
end
subplot(1,2,1),plot(1:700,V0*ones(1,700),’r.’,1:700,V(1:700),’k-’)
xlabel(’Time’), ylabel(’Volatility’),
legend(’Mean volatility’, ’Stochastic volatility’)
%Autocorrelation function
maxlag=10;
AC=zeros(1,maxlag);
for k=1:maxlag
AC(k)=(sum(V(1:N-k).*V(k:N-1))/(N-k-1)-mean(V)^2)/V0^2;
end

74
 subplot(1,2,2), semilogy(AC)
xlabel(’Time lag’),ylabel(’Autocorrelation’)
%%%%%%%%%%%%%%%%%%%%%%

Simulation results for ARCH(1) with γ = 0.4 and V0 = 1 are presented

in Fig. 3.6. One can observe both clamping and mean reversion in the
sequence of volatilities. Unfortunately, the autocorrelation function exhibits
the exponential decay, which contradicts the observed data, see Fig. 3.6.
1
12 10
Mean volatility
Stochastic volatility

10

0
10

8
Autocorrelation
Volatility

−1
6 10

−2
10

−3
0 10
0 200 400 600 800 0 2 4 6 8 10
Time Time lag

Figure 3.4: Volatility time series and autocorrelation function for ARCH(1).

Currently there exist more advanced models capable of much more accu-
rate reproduction of the statistics of price fluctuations, for example, general-
ized ARCH or GARCH(p,q) (Bollerslev (1986)):

D(n) = σ(n)W (n), (3.6.11)

q p
X X
2 2
σ (n) = γV0 + αk D (n − k) + βk σ 2 (n − k). (3.6.12)
k=1 k=1

Note than in the GARCH model, the future volatility depends not only on
price increments but also on the past volatilities. As a result, the GARCH

75
time series tends to have a longer memory and a stronger clamping compared
with ARCH model.
By now we are convinced that stochastic volatility in ARCH-type mod-
els exhibits the correct qualitative behavior. But what are the implications
for the statistics of price increments? For instance, can it be close to Gaus-
sian thus justifying a Black-Scholes type theory? The standard measure of
closeness of a random variable X to Gaussian is its kurtosis,

E((X − E(X))4 )
KX = − 3.
E((X − E(X))2 )2

If X is Gaussian, its kurtosis is equal to zero. It is easy to compute kurtosis

of Dn for n → ∞:
E(Dn4 ) = E(σn4 Wn4 ) = 3E(σn4 )
2
= 3E((γV0 + (1 − γ)Dn−1 )2 ) = 3γ 2 V02 + 6(1 − γ)γV02 + 3(1 − γ)2 E(Dn−1
4
).
Therefore, we get a difference equation for the fourth moment E(Dn4 ). The
solution converges to a limit as n → ∞ if 3(1 − γ)2 < 1. The limiting value
is
2γ − γ 2
lim (E(Dn4 )) = 3V02 .
n→∞ 1 − 3(1 − γ)2
The kurtosis is
(1 − γ)2
KD ∞ = 6 > 0.
1 − 3(1 − γ)2
Positive kurtosis suggests tails of the probability distribution ’fatter’
√ than
Gaussian. Moreover, kurtosis approaches infinity as γ → 1 − 1/ 3. This
suggests polynomial tails of the corresponding probability distribution (Try
to think about it in terms of Chebyshev inequality). In fact we have the
following

Conjecture:
Const
lim P rob.(Dn2 > V ) ∼ (1 + O(1/V )) ,
n→∞ Vµ
where exponent µ > 0 solves the equation

Γ(µ + 1/2)
√ = (2(1 − γ))−µ , (3.6.13)
π

76
where Γ(x) is gamma function.
To the best of my knowledge, the rigorous proof of this conjecture is an
open problem, but an informal argument leading to the above equation is
pretty straightforward. √
What is the exponent µ at the point γ = 1 − 1/ 3 for which kurtosis
becomes infinite? Using the fact that
3√
Γ(5/2) = π,
4
we find that µ = 2. As we know, the random variable distributed according
to P (V ) ∼ V −2 violates assumptions of CLT and the corresponding sample
mean does not converge to a Gaussian distribution. The fact that volatility
modeling led us to distributions of price fluctuations with power law tails has
fundamental implications for risk estimation.10

3.6.1 Checking the algebraic tails conjecture numeri-

cally.
The calculation which leads to (3.6.13) is quite involved and not necessarily
revealing. It is however instructive to try and confirm it numerically. One
way of deciding whether the tail of P rob(Dn2 > V ) follows a power law V −µ is
to plot the corresponding probability distribution function on a log-log plot:
the p. d. f. ̺(V ) ∼ V −µ−1 will look on this plot as a straight line with
the slope (−µ − 1). The MATLAB file creating such a p. d. f. numerically
is shown below. Note the application of the command f solve for numeric
solution of (3.6.13).
%%%%%%%%%%%%%%%%%%%%%%
%Numerical study of pdf tails in ARCH(1)
%%%%%%%%%%%%%%%%%%%%%%
%Expected variance
V0=1;
%Runlength
N=100;
%The number of experiments
expers=1000000;
10
See Alexander J. McNeil, Extreme Value Theory for Risk Managers for more informa-
tion.

77
%Final increments
Dfinal=zeros(1,expers);

for pp=1:expers

%Variance array
V=V0*ones(1,N);

%Increments array
D=zeros(1,N);

%Memory parameter
gamma=0.1;

for k=2:N

V(k)=gamma*V0+(1-gamma)*D(k-1)^2;
D(k)=sqrt(V(k))*randn;

end Dfinal(pp)=D(N)^2; end

range=[0:.1:100*V0];

pdf_exp=hist(Dfinal,range);

%Theory
%Analytical value of \mu
x=fsolve(’(2*(0.9))^x*Gamma(x+1/2)/sqrt(pi)-1’,2,optimset(’fsolve’))

loglog(range,pdf_exp/sum(pdf_exp),’k--’,range,range.^(-x-1),’r-’)
legend(’Experiment’,’Theory’)

The result for various values of the parameter γ is shown in Figs. 3.5, 3.6
and 3.7. As one can see, numerical simulations support both the algebraic
tail conjecture and the particular value of the tail exponent given by (3.6.13).

78
 3
10
Experiment
Theory
2
10

1
10

0
10

−1
10

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10
−1 0 1 2
10 10 10 10

Figure 3.5: The p. d. f. of Dn2 , γ = 1 − √1

3

3.6.2 MATLAB exercises for week 7.

• Finish the investigation of the probability of loss using the theory of large devi-
ations (see Week 6 exercises.)

• Advanced problem: Investigating the probability of being ahead. (To

be attempted only by students who have completed their previous assignment
early.) Suppose that two players flip a coin 1000 times each. The head brings
one point to the first player, the tail - one point to the second player. Investigate
the following question using MATLAB:

What is the probability P+ that either of the players has been ahead in this game
96 percent of the time? What is the probability P0 that neither player has been
ahead more than 52 percent of the time? How do these probability compare?

Comment. You can set your investigation up as follows. Let ξi = 1 if player I

79
 3
10
Experiment
Theory
2
10

1
10

0
10

−1
10

−2
10

−3
10

−4
10

−5
10

−6
10
−1 0 1 2
10 10 10 10

Figure 3.6: The p. d. f. of Dn2 , γ = 0.2

1
Pn
wins, and ξi = −1 if player II wins in the i-th round. Let Yn := n i=1 ξi . Then
#{0 ≤ n ≤ 1000; Yn > 0}
P+ = P rob( ≥ 0.96)
1000
#{0 ≤ n ≤ 1000; Yn > 0}
+P rob( ≤ 0.04)
1000
#{0 ≤ n ≤ 1000; Yn > 0}
P0 = P rob(0.48 ≤ ≤ 0.52)
1000
A naive argument aiming to compare P+ and P0 goes as follows: for large n, Yn
are mean zero variables with an approximate distribution N (0, 2√1N ). Therefore,
Yn is positive for about 50 percent of the time. Therefore, P0 >> P+ .
This argument misses completely the fact that random variables Y1 , Y2 , . . . are
correlated: If for example, Yn > 1, then Yn+1 > 0 with probability one.

Hints. Useful MATLAB commands: cumsum, sum(Y > 0, 2), sum(Z <
520). ∗ (Z > 480), 2). According to my estimates, P+ ≈ 0.25, P0 ≈ 0.023.
Hence P+ >> P0 .

80
 γ=0.1
1
10
Experiment
Theory
0
10

−1
10

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10
−1 0 1 2
10 10 10 10

Figure 3.7: The p. d. f. of Dn2 , γ = 0.1

3.7 The Gambler’s Ruin Problem

In finance one often has to deal with Wiener processes which hit boundaries.
A simple example is served by barrier options.11
Let us now look at a simple discrete version of this.
Assume a gambler increases his/her fortune by 1 with probability p and
loses 1 with probability q = 1 − p (the analogue for stocks is the logarithm of
the price). The gambler has initially fortune z (a non-negative integer) and
the total fortune inside the game is a ∈ N. We assume that 0 ≤ z ≤ a. The
player has fortune z(t) at time t ∈ N and the game stops when z(t) = 0 or
z(t) = a is achieved for the first time (either the gambler is ruined, or has
won everything). Let Ωz be the set of random sequences (z0 = z, z1 , z2 , . . . )
with zi ∈ {0, 1, . . . , a}, i ≥ 0 such that:
11
For example, a European up-and-out option call option may be written on an
underlying with spot price of $100, and a knockout barrier of $120. This option behaves
in every way like a vanilla European call, except if the spot price ever moves above $120,
the option ”knocks out” and the contract is null and void. Note that the option does
not reactivate if the spot price falls below $120 again. Once it is out, it’s out for good.
(Wikipedia)

81
 • If zi ∈ (0, a), zi+1 − zi = 1 with probability p and zi+1 − zi = −1 with probability
q = 1 − p.

• If zi = 0 or zi = a, zi+1 = zi with probability 1.

• Each sequence contains finitely many zi ’s not equal to 0 or a. Therefore, sequences

wandering between 0 and a forever without hitting either boundary do not belong
to Ωz .

We can visualize an element of Ωz as the trajectory or path of a one-

dimensional random walker in the strip {0, 1, . . . , a} × {0, 1, 2, . . .} of (z, t)
plane originating from (z, 0), see Fig. 3.8.

Symmetric random walk on the interval with sticky boundaries

10

7
Walker‘s co−ordinate

0
0 5 10 15 20 25 30
Time

Figure 3.8: The representation of a gambler’s fortune for p = 1/2 as a sym-

metric random walk in the interval with absorbing boundaries.

Perfectly sticky boundaries 0 and a of the interval containing gambler’s

fortune are often referred to as absorbing.
The probability measure on Ωz is easy to define: Each path belonging
to Ωz contains finitely many transitions zi − zi−1 ∈ {±}. Therefore, the
probability of a given path ωz = (z0 = z, z1 , z2 , . . . ) is p(ωz ) = pk q l where k

82
is the number of transitions zi+1 − zi = 1 (for i ≥ 0) and l the number of
transitions zi+1 − zi = −1 in this path.
The above definition of the measure on Ωz is simple as we chose to deal
only with the paths which ’stick’ in finite time. But there is a price to pay:
we will have to show that Ωz with the described probability measure is indeed
a probability space, i. e. that
X
Pz ≡ p(ωz ) = 1
ωz ∈Ωz

Keeping in mind that we will have to repay this debt, let us proceed to
the computation of the probability of winning. Let
′
X ′′
X
Wz = p(ωz ), Lz = p(ωz ),
P′
P′′ is summation over all ’winning’ paths ωz ∈ Ωz which end up at a
where
and is the summation over all ’losing’ paths which end up at 0. Clearly,
Wz is the probability of winning, Lz is the probability of losing and

Pz = Wz + Lz .

As functions of z, these probabilities satisfy the following boundary con-

ditions
P0 = 1, W0 = 0, L0 = 1,
Pa = 1, Wa = 1, La = 0.
In addition, the probabilities Wz , Lz and Pz satisfy some second order
difference equations, which we will now derive. Let us fix 0 < z < a. Then

Wz = p · P rob(W in | First jump is up) + q · P rob(W in | First jump is down)

= pWz+1 + qWz−1 .
The equations for Pz and Lz can be derived in exactly the same way:

Pz = pPz+1 + qPz−1 , 0 < z < a,

Lz = pLz+1 + qLz−1 , 0 < z < a.

83
Proposition 3.7.1. There is a unique solution to the second order homoge-
neous linear difference equation

Az = pAz+1 + qAz−1 , p + q = 1, p, q > 0, (3.7.14)

subject to the boundary condition

A 0 = C1 , A a = C2 ,

given by
λz − 1
Az = C1 + (C2 − C1 ), λ 6= 1,
λa − 1
z
Az = C1 + (C2 − C1 ), λ = 1,
a
where
q
λ=
p
Proof: We will give a constructive proof of the statement, the ideas from
which can be applied to the study of a large class of difference equations.
Let
DAk = Ak+1 − Ak
be the discrete derivative. The repeated application of discrete differentiation
gives
D2 Ak = Ak+2 − 2Ak+1 + Ak .
Using the operations of D and D2 , one can re-write equation (3.7.14) as
follows:
q−p
D2 Ak = DAk .
p
Let us introduce a new dependent variable Bk = DAk . In terms of this
variable our equation takes the form
q−p
Bk+1 = (1 + )Bk ≡ λBk , k = 0, 1, . . . , a − 1.
p
This equation can be solved explicitly:

B k = λk B 0 .

84
The solution is manifestly unique given B0 . Matrix counterparts of the equa-
tion we have just solved play fundamental role in the general theory of dif-
ference equation. Reverting to the original variable Ak we find that

Ak+1 = Ak + λk B0 .

The above recursion also has a unique solution given A0 :

k−1
X
Ak = A0 + B0 λm .
m=0

λk −1
The sum of the geometric series above is equal λ−1
for λ 6= 1. For λ = 1 it
is equal to k. Therefore,
λk − 1
Ak = A0 + B0 , λ 6= 1, (∗)
λ−1
Ak = A0 + B0 · k, λ = 1.
The constructed solution is unique given A0 , B0 . If Aa is specified instead,
λa − 1
Aa = A0 + B0 ,
λ−1
which always has a unique solution with respect to B0 (treat λ = 1 case in
the sense of a limit):
Aa − A0
B0 = a (λ − 1).
λ −1
Substituting this into (∗) we get the statement of the proposition for A0 = C1 ,
Aa = C 2 .

We can now solve the equations for Pz , Wz and Lz with their respective
initial conditions. Substituting C1 = C2 = 1 into 3.7.14 we find that

Pz = 1, ∀z.

Therefore, p(ωz ) does indeed define a probability measure on Ωz .

The computation of Wz and Lz is equally easy given (3.7.14):
First case: λ 6= 1 (p 6= q).
1 − (q/p)z
Wz = ,
1 − (q/p)a

85
 L z = 1 − Wz .

Second case: λ = 1 (q = p = 1/2).

z
Wz = ,
a
a−z
Lz = 1 − Wz = .
a

Note that in both cases, the chance of winning increases with the initial for-
tune of the gambler. Perhaps this is the reason that Keynes said ”Millionaires
should always gamble, poor men never”.12
The idea of solving equations in order to determine various statistical
properties of a random process is a powerful one. For instance, we can
easily calculate the expected duration (number of rounds E(τz ) of the game
started at z. (Note that the expected duration of the game is finite: the
game definitely ends when we have a run of a consecutive wins or losses.
The probability that this happens within a run of a games is more than
pa + q a > 0. Therefore, the probability that the game ends before time n · a
is ≥ 1 − (1 − (pa + q a ))n . This means that PP
{τ (ωz ) ≥ n} goes exponentially
fast to zero and that (Duration = Ez (τ ) := τ (ωz )p(ωz ) < ∞.)
The expected duration of the game satisfies the following difference equa-
tion

E(τz ) = p · E(τz | First round won) + q · E(τz | First round lost)

= p · (E(τz+1 ) + 1) + q · (E(τz−1 ) + 1)
= pE(τz+1 ) + qE(τz−1 ) + 1, 0 < z < a,
which should be equipped with the obvious boundary conditions:

E(τ0 ) = E(τa ) = 0.
12
John Maynard Keynes, 1st Baron Keynes, CB (pronounced keinz) (5 June 1883 - 21
April 1946) was a British economist whose ideas, called Keynesian economics, had a major
impact on modern economic and political theory as well as on many governments’ fiscal
policies. He advocated interventionist government policy, by which the government would
use fiscal and monetary measures to mitigate the adverse effects of economic recessions,
depressions and booms. He is one of the fathers of modern theoretical macroeconomics.
He is known by many for his refrain ”In the long run, we are all dead.” (Wikipedia)

86
 The difference equation we are dealing with is inhomogeneous. The sim-
plest way to solve it, is by guessing a particular solution. For example, for
p 6= q, a good guess would be

Fz = Cz,
1
where C is a constant. Substituting Cz into the equation we find C = q−p
.
Then
z
E(τz ) = + Az ,
q−p
where Az solves the homogeneous equation (3.7.14) subject to the boundary
conditions
a
A0 = 0, Aa = − .
q−p
Using the proposition (3.7.1), we find that

1 − (q/p)z
  
1
E(τz ) = z−a
q−p 1 − (q/p)a

Similarly, if p = q = 1/2, a reasonable guess for the particular solution is

2
Cz (second derivative of the quadratic function is constant). Substituting
this ansatz into the equation for the expected duration we find that C = −1.
Solving the corresponding homogeneous equation we get

E(τz ) = z(a − z).

This answer agrees with our intuition: if p = q = 1/2, random walk is
symmetric and the expected duration is maximal for z = a/2. Note also that
the answer for p = q = 1/2 exhibits a typical scaling: the expected time to
hit the boundary scales quadratically with the distance to the boundary.
The table below collects some numerical conclusions about the random
walk with adsorbing boundaries derived using the formulae obtained above.13
It is curious to notice that even if the odds are against you (say p = 0.45
or p = 0.4) you still have a good chance of winning if your initial capital is
comparable to the total capital of the game.
13
W. Feller, An introduction to probability theory and its applications

87
 p q z a Ruin Success E(Gain) E(Duration)
0.5 0.5 9 10 0.1 0.9 0 9
0.5 0.5 90 100 0.1 0.9 0 900
0.5 0.5 900 1000 0.1 0.9 0 90,000
0.5 0.5 950 1000 0.05 0.95 0 47,500
0.5 0.5 8,000 10,000 0.2 0.8 0 16,000,000
0.45 0.55 9 10 0.210 0.79 -1.1 11
0.45 0.55 90 100 0.866 0.134 -76.6 765.6
0.4 0.6 90 100 0.983 0.017 -88.3 441.3
0.4 0.6 99 100 0.333 0.667 -32.3 161.7

3.7.1 Nidhi’s game.

Imagine that you are already employed by a bank or another very respectable
financial institution. It is Friday evening and your manager invites you to
play a simple game we described above: with probability p you win, with
probability q your manager wins. The little catch is that you’re playing
using the content of your wallet, whereas your manager is playing using the
resources of the whole bank. Should you accept the offer?
Nidhi’s game is the a → ∞ version of the game we analyzed above.
Therefore, for p > q the probability of losing
 z
q
Lz ≈
p
decays exponentially with z. Unfortunately, the expected time of the game
diverges linearly with a:
a
E(τz ) ≤ .
p−q
If the game is ’fair’ and p = q = 1/2, you will lose with probability close
to one:
z
Lz = 1 − .
a
But at least you’ll have plenty of time for a long nice chat with the manager,
as the expected duration of the game is large:
E(τz ) ≈ az.
If q > p, you will lose with probability exponentially close to one and in
finite time. Given that gambling is addictive and a casino’s fortune a = ∞
for all practical purposes, here is a sound mathematical advice: don’t gamble.

88
3.8 Cox-Ross-Rubinstein model and Black-
Scholes pricing formula for European op-
tions.
We will now apply our knowledge of discrete random walks to analyse price
fluctuations of assets in Cox-Ross-Rubinstein (CRR) model and apply the
result to derive one of the most fundamental formulae of the quantitative
finance - Black-Scholes formula for the price of a European option.

3.8.1 The model.

Assume that we follow the price of an asset (e. g. a Microsoft share) for
the period of time T . Assume that we check the price at intervals ∆T . As
T
a result we perform N = ∆T price checks. Let Sn be the price at the n-th
check. According to CRR model which we discussed briefly in Chapter 1,

uSn with probability p

Sn+1 = {
dSn with probability q = 1 − p.

Here u > 1 is the price increase factor, 0 < d < 1 is price reduction factor,
p is the probability of price increase. In CRR model it is assumed that the
magnitude of the upward jump is the same as the magnitude of downward
jump,
u · d = 1.
Let parameterise u, d as follows:

u = 1/d = eσ ,

where σ > 0 is a parameter equal to the root mean square of the logarithmic
increment: v
u  2 !
u
tE ln S n+1
= σ.
Sn

For p = 1/2, σ has the meaning of volatility - the standard deviation of

logarithmic price increments. CRR model is designed to describe the con-
tinuous limit when ∆T → 0. As we will see below, p approaches 1/2 in the
continuous limit. We will therefore always refer to σ as volatility

89
 The CRR model can be defined more compactly, if one introduces a se-
quence of i. i. d. random variables ξn , which take values 1 and −1 with
probability p and q correspondingly. Then
Sn+1 = eσξn Sn .
Note that
ln(Sn+1 ) = ln(Sn ) + Dn ,
where Dn = σn ξn is the logarithmic price increment. Therefore the CRR
model makes the standard assumption about statistical independence of log-
arithmic price increments, which makes the prediction of price increments
impossible. We see that the sequence lnS0 , lnS1 , lnS2 , . . . describes the tra-
jectory of a one dimensional random walker with jumps of size σ.
CRR model has two parameters - σ and p. As we said, parameter σ
is related to price volatility. Amazingly, parameter p can be expressed in
terms of σ using the most fundamental assumption about the real markets -
the risk neutrality. According to this assumption, the expected value of an
asset T days (months, years) from now is equal to the risk free return one
would get by cashing the asset in today and putting money in the bank for T
days. Without some form of risk neutrality, the stable market is impossible:
imagine for example, that the expected return from an asset is less than
the risk-free investment. Then everybody would stop buying shares and take
their money to the bank. The trouble is, the bank wouldn’t want to invest
either!
For the case at hand, the risk-neutrality assumption takes a simple form:
er∆T Sn = E(Sn+1 | Sn ), (3.8.15)
where r is the interest rate which we assume to be constant. As we can see,
the interest is measured in units of inverse time.
Explicitly,
er∆T = up + d(1 − p),
which can be solved for p:
er∆T − e−σ
p= .
eσ − e−σ
Risk neutrality assumption even yields an important restriction on σ: as p is
the probability, we must have p < 1. Therefore
σ > r∆T

90
If in particular, we reduce intervals between the price measurements, ∆T →
0, σ cannot go to zero faster than ∆T without the model violating the risk
neutrality condition.
The price fluctuations in the CRR model are completely characterized:
the price logarithm follows a discrete random walk with jump size equal
to volatility. Jump probability is a function of interest rate and volatility.
Volatility has to be found from the analysis of market data. Note however,
that volatility in CRR model is constant, which seriously reduces the domain
of its applicability.
How can we use the information about asset prices to calculate a ’fair’
option value in the CRR Universe? To be definite, consider a European call
option. As we briefly discussed before, this is a contract, which gives its
owner a discretion to buy the specified asset (underlying asset) from the
contract’s writer at the specified time T (expiry or maturity date) for the
specified amount K (strike price). The writer has an obligation to sell the
underlying asset at the strike price.
Let S(T ) be the asset’s price at time T . If S(T ) > K, it makes sense for
the owner of the option to exercise it: by buying the option for K, he/she
can immediately sell it for S(T ) thus making a profit of S(T ) − K. If on the
other hand, S(T ) < K, the option is worth nothing as exercising it will lead
to a loss. Therefore, the value of European option at the exercise date is
C(N, SN ) = max(SN − K, 0),
where we used discrete time N = T /∆T . Note C(N, SN ) is a random vari-
able. The ’fair’ value of the option is fixed by the same risk-neutrality con-
ditions which we used to calculate the parameters of the CRR model. Let
C(n, S) be the value of the option on day n, S is the corresponding price of
the underlying asset. It is important to notice that in the CRR Universe,
the value of the option can depend only on S and n, as part of the price
trajectory for t > n is independent of the past walk history conditional on
the price Sn = S. If we sell the option on day n and put all the money in
the bank at the interest rate r, we will end up with risk-free pay-off equal to
C(n, S)e(N −n)r∆T
The risk neutrality principle requires that this is equal to the expected pay-off
resulting from exercising the option,
E(max(SN − K, 0) | Sn = S).

91
Therefore, we arrive at the formula for the value of the option:

C(n, S) = e(n−N )r∆T E(max(SN − K, 0) | Sn = S). (3.8.16)

The random variable inside the expectation value depends only on the num-
ber of up and down jumps between times n and N . In principle, we could use
this fact to calculate this expectation value using a version of central limit
theorem.
There is however, a more instructive way. We have learned recently that
one can compute various expectation values and probabilities for random
walks by solving difference equations. Can we derive an equation for C(n, S)?
Certainly:

C(n, S)
(n−N )r∆T
=e (p · E(max(SN − K, 0) | Sn = S, ξn = 1)
+q · E(max(SN − K, 0) | Sn = S, ξn = −1))
(n−N )r∆T
=e (p · E(max(SN − K, 0) | Sn+1 = uS)
+q · E(max(SN − K, 0) | Sn+1 = dS)).

Expressing the right hand side in terms of C(n, S) we find the equation

C(n, S) = e−r∆T (pC(n + 1, Su) + qC(n + 1, Sd)) , (3.8.17)

the right hand side of which is very similar to time-independent difference

equations we have been studying in the previous Section. Equation (3.8.17)
is the discrete counterpart of the celebrated Black-Scholes equation.
To be able to extract option price from the above equation we also need
boundary conditions. If at some time the value of the asset becomes zero, it
will stay zero forever. Therefore, the (3.8.16) gives

C(n, 0) = 0.

If S becomes very large at some time it is likely to stay very large up to time
N . The expectation value in the pricing formula becomes approximately
equal to S and we get the second boundary condition:
 
C(n, S)
lim = e(n−N )r∆T
S→∞ S

92
Finally, the value of the option at n = N is known:

C(N, S) = max(S − K, 0). (3.8.18)

To find the value of the option on any day n < N all we need to solve the
Black-Scholes equation backward in time using the ’final’ condition at n = N
we have just derived.14

3.8.2 Solving the discrete BS equation using binomial

trees.
It is very easy to solve (3.8.17) numerically using the binomial trees which
we will now describe. Suppose that we know the price of the underlying
asset S at time 0. We would like to know the value C(S, 0) of the European
call option based on this asset. Assume that the exercise time is T = ∆T N .
Equation (3.8.18) gives us the value of the option for any price of the under-
lying asset at time T which is N jumps away from time 0.
C(S, 0) can be computed in two steps:

(i) Construction of the price tree. Firstly, we should build the set of all
possible asset prices at times ∆T, 2∆T, . . . , N ∆T . For instance, at time ∆T
possible prices are uS and dS, at time 2∆T the prices are u2 S, udS = S, d2 S.
In general, at time n∆T (n ≥ 0) there are n + 1 possible prices: un−2m S,
m = 0, 1, . . . , n.
It is due to condition ud = 1 that the number of vertices grows lin-
early with time rather than exponentially. Such a binomial tree is called
recombining.
14
A practitioner’s view of the CRR model: Considered as a model CRR isn’t too pro-
ductive since it does not admit explicit solutions (other than as series or in the BS limit),
it doesn’t readily generalise, and empirically it falls over straight away The world is cer-
tainly not Wiener (whatever else it might be) but it is even less binomial! As a numerical
method it has a greater appeal.
CRR came along in the early 70s (CR wrote a book about BS that introduced a binomial
tree). It was seen at the time as being a simple (the marketing blurb) method for (i) pricing
American option in BS framework and (ii) explaining to non-maths types how you could
get prices out of BS. There were derivations around back then that extracted BS from
CRR. Of course CRR can be viewed as a weak numerical scheme for GBM, as can Jarrow
and Rudd and the other binomial methods that were circulating at the time. (Pr. Nick
Webber, WBS.)

93
 This set of prices can be thought of as the set of vertex labels of binomial
price tree, see Fig. 3.9, part I. The tree can be constructed recursively using
the f orward step shown in Fig. 3.9, part II. The vertices of the tree can
be arranged into levels according to the number of time steps needed to get
from the root of the tree to the given vertex. Vertices at the final, N -th level,
are called leaves.

(ii) Calculation of the option price. Leaves of the tree correspond to the
exercise time. Therefore, the value of the option at each leaf is known due to
(3.8.18). Assume that value of the option at each vertex at level n is known.
Then its value at each vertex of level n − 1 can be calculated using equation
(3.8.17)! This process is called backward evaluation. Its elementary step is
shown in Fig. 3.9, part III. After N steps of backward evaluation we reach
level zero, i. e. calculate C(S, 0).
The binomial tree method of solving the discrete BS equation can be gen-
eralized to deal with stochastic volatility models and/or more complicated
type of options such as American options. You will have a chance to im-
plement this algorithm and its generalizations during the MATLAB sessions
and while working on the term project.

3.8.3 The continuous limit of CRR model.

It is not difficult to solve the discrete Black-Scholes equation exactly, but we
will not do it. Instead we will calculate the option price in the limit ∆t → 0,
as this will motivate our future study of continuous-time random processes
and differential equations.
At this point we need an extra assumption, that as ∆t → 0, σ → 0. This
is a requirement necessary (but not sufficient) for the continuity of prices:
the smaller the interval between the measurements, the smaller the price
increment of the underlying asset should be. Thus σ should be a function of
∆T , but its exact form is not known to us yet. For small σ and ∆T ,

e−r∆t ≈ 1 − r∆t,
1
u ≈ 1 + σ + σ2,
2
1
d ≈ 1 − σ + σ2,
2

94
 uN S

u2 S
uN-2 S
uS

S
S
dS
d2 S
dN-2 S

Level 1 d NS

Level 2
Level N
I. The binomial price tree

Cn+1 (uS)
uSn

Sn

dSn
Cn+1 (dS)

II. Iterative tree construction III. Backward Recursion Step

Figure 3.9: The binomial tree algorithm.

2σp ≈ r∆T + σ − σ 2 /2
∂ ∂ σ2 ∂2
C(uS, t+∆t) ≈ C(S, t)+ C(S, T )∆T +S C(S, T )(σ+σ 2 /2)+ S 2 2 C(S, T )
∂t ∂S 2 ∂ S

95
 ∂ ∂ σ2 ∂2
C(dS, t+∆t) ≈ C(S, t)+ C(S, T )∆T +S C(S, T )(−σ+σ 2 /2)+ S 2 2 C(S, T )
∂t ∂S 2 ∂ S
Substituting these expansions into (3.8.17) and collecting the terms we get
the following relation:

∂C ∂C σ2 2 ∂ 2C
(S, t) + rS (S, t) + S (S, T ) − rC(S, t) + O(σ 3 /∆T ) = 0.
∂t ∂S 2∆T ∂ 2 S
(3.8.19)

The price of the option should not depend on ∆T as the latter goes to zero.
Therefore, two cases are possible:

Case I. lim∆T →0 (σ 2 /∆T ) = 0. Then the limiting equation is

∂C ∂C
(S, t) + rS (S, t) − rC(S, t) = 0.
∂t ∂S
This is the first order partial differential equation, which can be solved using
the method of characteristic. Given the boundary and the final conditions
the solution is

C(S, t) = max(S − er(t−T ) K, 0) (∗∗)

The meaning of this solution is clear: assume that the asset price grows
deterministically according to the risk-free interest rate. Then the price neu-
trality assumption applied to C(S, t) and the known pay-off at time T would
tell us that
er(T −t) C(S, t) = max(Ser(T −t) − K, 0),
which coincides with (∗∗). In other words, Case I corresponds to zero volatil-
ity limit for which it would be impossible to make or lose money using options.
The possibility that real markets fall into this Case has been thoroughly ruled
out experimentally.

Case II. lim∆T →0 (σ 2 /∆T ) = σ02 > 0. The limiting equation is the celebrated
Black-Scholes (BS) equation:

∂C ∂C σ02 2 ∂ 2 C
(S, t) + rS (S, t) + S 2 (S, T ) − rC(S, t) = 0.
∂t ∂S 2 ∂ S
(3.8.20)

96
The parameter σ02 has the meaning of the continuous time volatility, asso-
ciated with the geometric Brownian motion which describes the continuous
limit of the CRR model. We will learn more about this stochastic process in
the near future.
In the mean time, let us solve equation (4.2.15). Introducing new inde-
pendent variables
x = ln(S/K), τ = T − t
we can re-write the BS equation as the forward diffusion equation with the
drift term:
∂C σ 2 ∂C σ2 ∂ 2C
− (x, τ ) + (r − 0 ) (x, τ ) + 0 2 (x, τ ) − rC(x, τ ) = 0.
∂τ 2 ∂x 2 ∂ x
Next, we change the dependent variable:

C(x, τ ) = eαx+βτ C0 (x, τ )

where parameters α and β are to be chosen in such a way that C0 (x, t)

satisfies the canonical diffusion equation:
σ02 2
−∂τ C0 (x, τ ) + ∂ C0 (x, τ ) = 0
2 x
1 σ2
+(−β − r + α2 σ02 + (r − 0 )α)C0 (x, t)
2 2
2
σ
+(r − 0 + ασ02 )∂x C0 (x, t) = 0.
2
The condition that terms proportional to C0 and ∂x C0 vanish leads to the
system of two equations for α and β:
1 σ2
−β − r + α2 σ02 + (r − 0 )α = 0,
2 2
σ02
r− + ασ02 = 0.
2
The solution is
1 r
α= − 2,
2 σ0
σ02 1 r
β=− ( + 2 )2 .
2 2 σ
97
Given this choice, the equation for C0 becomes the familiar f orward diffusion
equation:
σ02 2
−∂τ C0 (x, τ ) + ∂ C0 (x, τ ) = 0. (3.8.21)
2 x
Due to our choice of temporal variable τ = T − t, the final condition (3.8.18)
becomes the initial condition for C0 :

C0 (x, τ = 0) = Kmax(e(1−α)x − e−αx , 0) (3.8.22)

The boundary conditions become

Cx ∼ eC± x , x ± ∞, (3.8.23)

where C± are constants depending on α.

The solution to equation (3.8.21) subject to boundary conditions (3.8.23)
and initial conditions (4.1.2) exists and is unique:
Z ∞
C( x, τ ) = dyG(τ, x − y)C0 (y, τ = 0), (3.8.24)
−∞

where s
2
1 − y 2
G(τ, y) = e τ 2σ0
.
2πτ σ02
is the Green function of the diffusion equation. Notice that Green’s function
coincides with the Gaussian probability distribution function with variance
τ σ02 , which describes (due to CLT) the law of the position of a discrete random
walker in the continuous limit.
Substituting the initial conditions into (3.8.24) we find:
Z ∞ Z ∞
(1−α)x
C0 (x, τ ) = S dyG(x − y, τ )e −K dyG(x − y, τ )e−αx .
0 0

The integrals in the above solution can be expressed in terms of error func-
tions. Substituting the answer for C0 into C = eαx+βτ and re-writing the
result in terms of the original variables S and t, we arrive at the Black-
Scholes formula for the European call option:
erf c(−d1 ) erf c(−d2 )
C(S, t) = S −K , (3.8.25)
2 2
98
where t < T ,
σ02
ln(S/K) + (T − t)(r + 2
)
d1 = p , (3.8.26)
2σ02 (T − t)
σ02
ln(S/K) + (T − t)(r − 2
)
d2 = p , (3.8.27)
2σ02 (T − t)
(3.8.28)

and ∞
2
Z
2
erf c(x) = √ dye−y
π x

is the complementary error function. Let us spend some time analyzing BS

formula. First of all, consider the limit t → T −0. In this limit d1,2 → +∞ for
S > K and d1,2 → −∞ for S < K. As erf c(+∞) = 0 and erf c(−∞) = 0,
we get
C(S, T − 0) = S − K, S > K,
C(S, T − 0) = 0, S < K,
in agreement with the value of the European call option at the exercise time.
As t approaches the exercise time, C(S, t) considered as a function of S/K
converges to the exercise value while staying above it, see the left pane of
Fig. 3.10. This is the price of the opportunity: For as long as t < T there
is a chance for the price of the underlying asset rising sharply in price. The
price of the opportunity is also reflected in the increase of the BS price of
the option with volatility, see the right pane of Fig. 3.10.
Our derivation of Black-Scholes formula in the continuous limit of CRR
model demonstrates the usefulness of differential equations for quantitative
finance. The properties of many financial objects more complicated than Eu-
ropean options (path dependent options such as barrier options, American
options, stochastic volatility models, etc.) can be analysed using differen-
tial equations - either analytically or numerically. The rest of the course
is dedicated to the study of differential equations and stochastic differential
equations.

99
 Increase of the option price with the contract length Increase of the option price with volatility.
0.5

0.5
0.45

0.45
0.4
0.4
0.35
0.35
0.3
0.3
C

C
0.25
0.25

0.2
0.2

0.15
0.15

0.1 0.1
Time−to−exercise
increases
0.05 Volatility increases
0.05

0
0.5 1 1.5 0.5 1 1.5
S/K S/K

Figure 3.10: The price of the opportunity in Black-Scholes Universe.

3.8.4 Matlab Exercises for Weeks 8, 9.

Solving discrete BS equation using binomial trees.
Write a Matlab programme to calculate the price of a European call option
option with CRR model. Compare your result with Black Scholes formula
(use blsprice command in MATLAB). Suggested parameter values: T = 1
(year), r = σ02 = 0.1 (1/year), K = 1. Remember, that volatility in CRR
model is σ 2 = σ02 ∆T , where ∆T is step size.

100