Lecture notes for STAT3006 / STATG017

Stochastic Methods in Finance

Part 3 Julian Herbert Department of Statistical Science, UCL 2010-2011

Contents
7 Calculus refreshers 7.1 Taylor series . . . . . . . . 7.2 Chain rule dierentiation . 7.3 Partial dierentiation . . . 7.4 Linear ordinary dierential 54 54 55 55 56

. . . . . . . . . . . . . . . . . . equations

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

8 Continuous-time stochastic processes for stock prices 57 8.1 Random walk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 8.2 Other processes and Markov property . . . . . . . . . . . . . . . . . . . 58 8.3 Taking limits of the random walk . . . . . . . . . . . . . . . . . . . . . 59 8.4 Brownian motion (Wiener process) . . . . . . . . . . . . . . . . . . . . 59 8.5 Denition of Brownian motion . . . . . . . . . . . . . . . . . . . . . . . 60 8.6 Generalised Brownian Motion process . . . . . . . . . . . . . . . . . . . 63 8.7 It o process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 8.8 A process for stock prices: the geometric Brownian motion . . . . . . . 65 8.9 Appendix: Brownian Motion as a limit of a discrete time random walk 67 9 Introduction to stochastic calculus and It os lemma 9.1 Ordinary and stochastic calculus . . . . . . . . . . . . . . . . . . . . . . 9.2 It os formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.1 Derivation of the SDE of a generalised Brownian motion . . . . 9.3.2 Solution of the SDE of a geometric (exponential) Brownian motion 9.3.3 Logarithm of stock prices . . . . . . . . . . . . . . . . . . . . . 9.3.4 Generic transformation of stock prices . . . . . . . . . . . . . . 9.3.5 Forward price . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4 Appendix: second order eects in the limit . . . . . . . . . . . . . . . . 9.5 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 69 71 73 73 74 75 75 75 76 77

Chapter 7 Calculus refreshers

These notes are intended to give you reminders of various important classical calculus results that we will use. The result (7.3) and the solution to equation (7.4) are particularly relevant. For more details see an introductory calculus book.

7.1

Taylor series
f (x0 ) f (x0 ) (x x0 )2 + (x x0 )3 + ... 2! 3!

A function f (x) can be expanded as a Taylor series around the point x0 as follows f (x) = f (x0 ) + f (x)(x x0 ) + =
f (i) (x0 )(x x0 )i i=0

i!

where f (i) is the ith derivative of f , and we need to assume that the derivatives at x0 all exist. The special case of this expansion in which x0 = 0 is sometimes called Maclaurins series. If f () is a function of two variables, say f (x, y ), then a Taylor series expansion can be made about a point (x0 , y0 ) as follows f (x, y ) = f (x0 , y0 ) + f (x0 , y0 ) f (x0 , y0 ) (x x0 ) + (y y0 ) + x y 1 2 f (x0 , y0 ) 1 2 f (x0 , y0 ) 2 ( x x ) + (y y0 )2 + 0 2 2 2! x 2! y 1 2 f (x0 , y0 ) 2 (x x0 )(y y0 ) + ... 2! xy

(7.1)

This can be generalised to the case where f is a function of n variables.

54

Stochastic Methods in Finance

7.2

Chain rule dierentiation

The classical chain rule for dierentiation says that for dierentiable functions f and g we have [f (g (s))] or equivalently df dg d f (g (s)) = (g (s)) (s) ds dg ds or also equivalently f (g (t)) f (g (0)) = =
0

= f (g (s)) g (s)

f (g (s)) g (s)ds f (g (s))dg (s)

where we have written h (t) for the ordinary derivative of a function h at t.

7.3

Partial dierentiation

For f a function of x and y , with x and y both functions of a single independent variable t, the extension of the chain rule gives us df f (x, y ) dx f (x, y ) dy = + dt x dt y dt We can also write this as df = f f dx + dy x y (7.2)

when y and x depend on one or more other variables. For our course it will be useful to consider this result heuristically as an application of the Taylor series. If we take the limit as x0 tends to x and y0 tends to y and we write dx x x0 , dy y y0 and df (x, y ) f (x, y ) f (x0 , y0 ), then we can express result (7.1) as f f dx + dy + x y ( ) 1 2f 2 2f 2 2f dx + 2 dy + 2 dxdy + ... 2! x2 y xy

df =

(7.3)

In ordinary calculus the second order (and higher order) terms in (7.3) tend to zero so that we obtain result (7.2).
55 J Herbert UCL 2010-11

Stochastic Methods in Finance

7.4

Linear ordinary dierential equations

A vast topic - here we only present one simple ordinary dierential equation (ODE) to remind the reader of one basic technique, and of the important example of exponential growth. The ODE dx = rx, dt (7.4)

(also sometimes written as dx = rx dt), where r is a constant, can be solved using the variable separable technique, or in other words by writing the ODE as the integral equation
1 dx = 1 dt rx

Integrating this then gives

1 ln(x) r

= t + k1 = rt + rk1 = ert+rk1 = Kert (7.5)

ln(x) x x

where k1 and K are constants. This is the equation for exponential growth at rate r. To obtain particular solutions, dierential equations are usually solved together with boundary conditions. In the above simple example, we can determine the value of the arbitrary constant K with the added information that say x(0) = x0 . Then substituting this into (7.5) we can nd K as x(0) = x0 = Ker0 = K so that the particular solution to the ODE (7.4) together with the boundary condition (7.6) is x = x0 ert . When we move into the world of partial dierential equations an initial boundary condition can specify the value of the function across a curve. However boundary conditions do not always have to be initial conditions, they can, for example, be terminal conditions, or other more complicated types of boundary conditions. (7.6)

56

J Herbert UCL 2010-11

Chapter 8 Continuous-time stochastic processes for stock prices

We now explore further suitable stochastic processes to describe the behaviour of stock prices. A typical example of stock price movements is given in the graph below, which shows the price of the UK FTA index from 1963-1992.

The binomial model we saw in the past lectures is one approach to modelling the movement of real prices. In order to implement it on realistic time scales we will need to have a large number of time steps and hence use a large number of nodes. Although there are sometimes benets to the discrete time approach to modelling the stock prices, solving for derivative prices can be computationally expensive as we are often not able to obtain explicit solutions over many nodes. One alternative approach that can lead to the derivation of some explicit theoretical results for option pricing is to consider a continuous time limit. 57

Stochastic Methods in Finance

A continuous time limit of the binomial model can lead us to a model based on the important stochastic process Brownian motion. This increases our exibility for asset price modelling and allows the use of mathematical techniques such as stochastic calculus. First of all we look again briey at a special case of the binomial model, and how it can be extended over time to model the stock prices. We will then consider the appropriate limits to use to move into continuous time.

8.1

Random walk

We will start now from this simple discrete-time, discrete-value (discrete-state space) stochastic process that has natural similarities with the binomial model we just looked at. Here {xt } is a random variable that begins at a known value x0 , and at time t = i jumps by a random variable i from xi1 , for i = 1, 2, 3, . . . So for example x1 = x0 + 1 , and x2 = x1 + 2 . Assume that i are independent, identically distributed random variables with i = with probability i = + with probability
1 2 1 2

for i = 1, 2, . . ., so that xt takes a jump at each time step of size , either up or down, 1 each with probability 2 . The jumps are independent, and therefore we can describe the dynamics of xt with the following equation xt = xt1 + t = x0 +
t i=1

. where t is a random variable that takes only values or with equal probability 1 2 t Therefore the random variable xt = x0 + i=1 i is the position after t steps of a random walker who started from x0 , where each step is equally likely to be up or down by . The random process {xt , t = 0, 1, . . . , N } is the path followed by the walker over time.

8.2

Other processes and Markov property

Because the probability of an up or down jump is 1 , at time t = 0 the expected value 2 of xt is x0 for all t. One way to generalise this process is by changing the probabilities for an up- or down-jump to be greater than 0.5 (random walk with drift). Another way to generalise this process is to let the size of the jump at each time t be a continuous random variable. For example, we could let the size of the jump be
58 J Herbert UCL 2010-11

Stochastic Methods in Finance

normally distributed with mean zero and variance 2 . In this case we will call xt a discrete-time, continuous-state stochastic process. The random walk (with discrete and continuous states, with or without drift) process satises the Markov property, and are therefore called Markov processes. This property is that the probability distribution for xt+1 depends only on xt , and not additionally on what happened before time t. For example, in the case of the simple random walk, if xt = 6 and = 1, then xt+1 1 . The values before xt are irrelevant once we can equal 5 or 7, each with probability 2 know xt .

8.3

Taking limits of the random walk

If our N steps take time T to be carried out, each one takes t = T /N time units. As we move from discrete to continuous time, both the time between jumps, t, and the size of the jumps, , will need to tend to zero. However the rate at which each tends to zero, and hence the relationship between the time gap t and the jump size , will also be important. We will need to x this relationship in some way when we take limits. We might initially think that as we take limits we should set = k t for some positive constant k , so that the ratio of /t is always constant. However it turns out that this choice will not work and we will need to nd another approach (see appendix). Remember that we are going to take limits as both t 0 and 0, whereas t is xed as the length of the overall interval we are looking at. It turns out that for the variance of the process to behave well and remain nite, it is the ratio 2 /t that we need to keep constant while t 0 and 0, rather than the ratio of /t. If we call this constant 2 := 2 /t, then we can see that the variance of an incremental change in the process over time t is 2 t, and hence is proportional to t, the size of the time length we are looking at. The standard deviation is therefore proportional to the square root of t. This property, that the variance of the incremental change in our process is proportional to the length of time we are considering, is one of the key properties of the continuous time stochastic process that we will use, Brownian motion.

8.4

Brownian motion (Wiener process)

Brownian motion or a Wiener process is a continuous-time stochastic process with the following important properties1 :
1

In 1827, the botanist Robert Brown rst observed and described the motion of small particles 59 J Herbert UCL 2010-11

Stochastic Methods in Finance

The Brownian motion process has independent increments. This means that the probability distribution for the change in the process over any time interval is independent of any other (non-overlapping) time interval (a property that is also true for the discrete time random walk). It is a Markov process. As explained earlier, this means that the probability distribution for all future values of the process depends only on its current value, and is unaected by past values of the process or by any other current information. As a result, the current value of the process is all one needs to make a best forecast of its future value. This property follows from the independent time interval property. Changes in the process over a nite interval of time are normally distributed, with a variance that increases linearly with the time interval. The Markov property is particularly important in the context of modelling stock markets, since it implies that only current information is useful for forecasting the future path of the process. Stock prices are often modelled as Markov processes, on the grounds that public information is quickly incorporated in the current price of the stock, so that the past pattern of the prices has no forecasting value. This is called the weak form of market eciency. If it did not hold, investors could in principle beat the market through technical analysis, by using the past pattern of prices to forecast the future. The fact that a Wiener process has independent increments means that we can think of it as a continuous-time version of a random walk.

8.5

Denition of Brownian motion

Brownian motion can be formally dened by the following properties. The process zt is Brownian motion if and only if 1. All non-overlapping increments of the process are independent (so that zs2 zs1 is independent of zt2 zt1 for all 0 s1 < s2 t1 < t2 ). 2. For 0 s < t the increment zt zs is normally distributed with mean 0 and variance t s. 3. zt is continuous and z0 = 0.
suspended in a liquid, resulting from the apparent successive and random impacts of neighbouring particles; hence the term Brownian motion. In 1905, Albert Einstein proposed a mathematical theory of Brownian motion, which was developed further and made more rigorous by Norbert Wiener in 1923.

60

J Herbert UCL 2010-11

Stochastic Methods in Finance

We can think of the second property as saying that a process z = {zt }t>0 following Brownian motion has increments that can be expressed as z = zt+t zt = t t where t N (0, 1). Let us examine what the conditions above imply for the change of z over some nite interval of time T . If we want to study the dierence zT z0 , we can use N intervals with each interval step being t = T /N , and obtain zT z0 = zT zT T /N + zT T /N zT 2T /N + + zT /N z0 N = i t
i=1

where the i N (0, 1) and they are independent. Each small dierence has mean 0 and variance equal to the interval length t, which for N intervals will be T /N . It follows that zT z0 is normally distributed with E [zT z0 ] = 0 and Var[zT z0 ] = N i.e. (zT z0 ) N (0, T ) T =T N

An example of a sample path of the Brownian motion process is shown below.

Figure 8.1: Example of a Brownian motion sample path Example. Suppose zt follows a Brownian motion process where z1 = 25 and t = 1 year. What are the distributions of z2 and z6 ? From above, we know that (zT2 zT1 )
61 J Herbert UCL 2010-11

Stochastic Methods in Finance

N (0, (T2 T1 )). Therefore for T2 = 2 and T1 = 1 we have (z2 25) N (0, 1) z1 N (25, 1)

Similarly, when T = 6 we have z6 N (25, 5). Notice that in the case where the process starts at zero, i.e. z0 = 0, the generic variable zt follows the normal distribution N (0, t). Example. If Z N (0, 1), then the process xt = tZ is continuous and is marginally distributed as a N (0, t). Is xt a Brownian motion process? The answer is no, since the increments respect the conditions for a Wiener ( do not ) process. In fact, we have xt+t xt = t + t t Z . It follows that the increment follows a normal distribution with zero mean and variance given by
(

)2 t t + t t 1 = t 2t 1 + 1 t

which is not t. Moreover, the increments are not independent. For instance, if we consider the two increments xt+t xt and xt x0 we have that their correlation is given by the following (where we use the fact that x0 = 0 and Z 2 is a 2 1 ): E (xt+t xt )(xt x0 ) = E (xt+t xt )xt = t + t tE (Z 2 ) E (x2 t) t = t 1 + 1 = 0 t It follows that xt is not a Brownian motion process. Example. If zt and wt are two independent Brownian motion processes starting at zero, and is a constant between 1 and 1, then the process xt = zt + 1 2 wt is continuous and has marginal distributions N (0, t). Is xt a Brownian motion? Here an increment is given by xt+t xt = (zt+t zt ) +

1 2 (wt+t wt )

We know from the properties of Brownian motion that both the increments of zt and of wt appearing above follow a N (0, t) distribution. It follows that the increment in xt is the sum of a N (0, 2 t) and a N (0, (1 2 )t), i.e. a N (0, t), which is consistent with the Brownian motion process properties. Moreover, the increment in xt shown above will be independent of any other increment in the process xt over a non-overlapping time interval, since this is true for both zt+t zt and wt+t wt . It follows that xt is indeed a Brownian motion.
62 J Herbert UCL 2010-11

Stochastic Methods in Finance

8.6

Generalised Brownian Motion process

If z is a Brownian motion (or equivalently Wiener) process , we have seen that E [z ] = 0 (drift) and Var[z ] = t (variance). We can construct a general class of processes x = {xt } such that for small time periods t x = xt+t xt = at + bz where a and b are constants. Then x N (at, b2 t) since z is normally distributed and: E [x] = at (new drift) and Var[x] = b2 t (new variance) With arguments similar to those in Section 1, we obtain that the behaviour of the change over a time interval T is given by (xT x0 ) N (aT, b2 T ) Example. Consider a generalised Wiener process with x0 = 50, a = 20 and b2 = 900. At the end of six months (T = 0.5) we have (x0.5 x0 ) N (0.5a, 0.5b2 ), and therefore the value of the variable after six months has distribution N (x0 + 0.5a, 0.5b2 ), i.e. a N (60, 450). We say that x satises the stochastic dierential equation dx = adt + bdz where a is the drift rate, b2 is the variance rate and dz is the error term. If b = 0 (no variability) then dx = adt, i.e. x = x0 + at: the process grows linearly with time. When b = 0, the term bz adds variability around the line x = x0 + at. Sample paths for a generalised Brownian motion process and Brownian motion are shown below.

8.7

It o process

A further type of stochastic processes can be dened where the drift and variance rate are not constant anymore. A random process x = {xt } is an It o process if for any t and very small t x = xt+t xt = a(x, t)t + b(x, t)z

63

J Herbert UCL 2010-11

Stochastic Methods in Finance

where a(x, t) is the drift rate and [b(x, t)]2 is the variance rate. We say that x satises the following stochastic dierential equation dx = a(x, t)dt + b(x, t)dz It o processes are sometimes called diusion processes, as they can be used to model the diusion of gas particles.

64

J Herbert UCL 2010-11

Stochastic Methods in Finance

8.8

A process for stock prices: the geometric Brownian motion

We are looking for an appropriate process to model stock prices. One model we may consider using for non-dividend paying stock prices is the generalised Brownian motion process dS = dt + dz where is the drift of the stock price and the , is the square root of the variance rate. The model implies that in a period of t S = St+t St = t + t where N (0, 1) as usual. Observe that the expected increase in the stock price in this time period is t, which is independent of the stock price itself. This does not seem appropriate because growth in a stock price is usually related to the size of the stock price itself. It is the return from a stock that we are interested in, which is measured in terms of a percentage change in price. This model will also allow the possibility of negative values for the stock price S which is clearly not appropriate. A more appropriate model is to look at the percentage change as following generalised Brownian motion, i.e. S St+t St = = t + z S St This gives a model for the actual stock price S as S = St t + St z i.e. S/S N (t, 2 t). We call this process a geometric Brownian motion, and we say that it satises the following stochastic dierential equation dS = Sdt + Sdz ()

If = 0 (no variance) then S is a risk-free asset and dS = Sdt, which is equivalent to dS/dt = S , i.e. S = S . The solution to this dierential equation is ST = S0 eT For = 0 the price grows at a continuously compounded rate of per unit. The geometric Brownian motion has two parameters: the rate of return and the volatility (risk).
65 J Herbert UCL 2010-11

Stochastic Methods in Finance

It turns out that when = 0 the process that solves the stochastic dierential equation () is given by
{(

St = S0 exp

2 t + z 2

(8.1)

i.e. () corresponds to the exponential of a generalised Brownian motion. We shall look into this further in lectures to come. Following the formula above, the geometric Brownian motion is also called exponential Brownian motion. Example. Consider a stock that pays no dividends, has a volatility of 30% per annum and provides an expected return of 15% per annum. Then = 0.15 and = 0.30: dS = 0.15dt + 0.30dz S This means that as an approximation we can write S = 0.15t + 0.30 t S If the price is now 100 what is the distribution of the price after one week? One week is 1/52 years, i.e. t = 0.0192; St = 100. We obtain: St+(1 week) 100 = 0.15 0.0192 + 0.30 0.0192 100 where N (0, 1). It follows: St+(1 week) = 100.288 + 4.16 Therefore our approximation gives: St+(1 week) N (100.288, 4.162 ) An approximate 95% probability interval is then given by 100.288 1.96 4.16 St+(1 week) 100.288 + 1.96 4.16, which gives 92.14 St+(1 week) 108.44. With the more precise lognormal formula for a 95% probability interval, derived from (8.1) above, Se( we obtain 92.17 St+(1 week) 108.49. They are almost identical, since t here is very small (0.0192). If t is larger the rst procedure is no longer accurate, and we need to use the lognormal approach.
66 J Herbert UCL 2010-11
2 )T 1.96 T 2

ST Se(

2 )T +1.96 T 2

Stochastic Methods in Finance

8.9

Appendix: Brownian Motion as a limit of a discrete time random walk

In later lectures we will be working with Brownian motion (also called a Wiener process), which is a continuous-time stochastic process. Here we provide an outline of how to move from a version of the discrete-time binomial model (the random walk) to continuous-time Brownian motion, by taking appropriate limits. Consider a process xn that follows the random walk outlined in section 8.1. After n steps the process can be expressed as xn = x0 +
t i=1

i .

We are interested in the incremental change xN x0 (i.e. the distribution of the increase or decrease in the process after N time steps) when N is large. Suppose now that the time between the jumps up or down of the random walk is small and is t, so that jumps take place at times t, 2t, 3t . . . The number of jumps up to a time point t will be n = t/t. As we move from discrete to continuous time, both the time between jumps, t, and the size of the jumps, , will need to tend to zero. However the rate at which each tends to zero, and hence the relationship between the time gap t and the jump size , will also be important. We will need to x this relationship in some way when we take limits. We might initially think that as we take limits we should set = k t for some positive constant k , so that the ratio of /t is always constant. However as we shall see, this choice will not work and we will need to nd another approach. As we know that n, the number of jumps between time 0 and time t, is going to become large in the limit, we can start by using the central limit theorem. One version of the central limit theorem tells us that if X1 , X2 , . . . , Xn are a sequence of independent identically distributed random variables with nite means and nite non-zero variances 2 , and Sn = n i=1 Xi , then Sn n N(0, 1) n
1 2 = 0), in distribution as n . In our case the mean of each jump i is zero ( = 1 2 and we can calculate the variance as

2 1 = 2( 2 E[2 +1 )0 i ] (E[i ]) 2

= 2.

67

J Herbert UCL 2010-11

Stochastic Methods in Finance

2 Therefore, the variance of n i=1 i is n as all covariances are zero by independence of the i . So for large n, we have that approximately n i=1

i N(0, n 2 ).

Substituting n = t/t as the number of jumps in our time interval for given t, the variance of the incremental change in the process, xt x0 , is 2 t. t Remember that we are going to take limits as both t 0 and 0, whereas t is xed as the length of the overall interval we are looking at. We can see from this variance result that, in order for the variance of the process to behave well and remain nite, it is the ratio 2 /t that we need to keep constant while t 0 and 0, rather than the ratio of /t. If we call this constant 2 := 2 /t, then we can see that the variance of an incremental change in the process over time t is 2 t, and hence is proportional to t, the size of the time length we are looking at. The standard deviation is therefore proportional to the square root of t. This property, that the variance of the incremental change in our process is proportional to the length of time we are considering, is one of the key properties of the continuous time stochastic process that we will use, Brownian motion.

68

J Herbert UCL 2010-11

Chapter 9 Introduction to stochastic calculus and It os lemma

In the last few lectures we have tried to model the behaviour of stock prices using some specic stochastic processes. But what about the behaviour of the price of a derivative? We expect that in general the price of a derivative will be a function of both time and the price of the underlying asset. If we specify a particular process for the price of the underlying asset, can we obtain the dynamic of the price of the derivative? The answer is yes, and this can be achieved using stochastic calculus.

9.1

Ordinary and stochastic calculus

In previous lectures we have seen that the various stochastic processes we analysed satised certain stochastic dierential equations (SDE). The most general form of SDE was of the type dxt = a(xt , t)dt + b(xt , t)dzt where {zt } is a standard Brownian motion, and a and b are some known (non-random) functions of xt and time. In fact, when we write an SDE we are really just using convenient notation to describe an integral equation. If we add an initial condition, say that the process starts at a given point x0 , then the SDE above can be interpreted as a version of the integral equation

t 0

x t = x0 +

a(xs , s)ds +

b(xs , s)dzs

(9.1)

Example. In the case of constant a(xt , t) = a and b(xt , t) = b we obtain the SDE of the generalised Brownian motion dxt = adt + bdzt 69

Stochastic Methods in Finance

and the associated integral equation simplies to xt = x0 + at + bzt We are of course comfortable with the integral with respect to t (or equivalently with respect to s in (9.1)), but we need to be careful with what it really means to talk about the second integral in (9.1), which is with respect to Brownian motion (zt ). We have seen in the lectures on Brownian motion that in the case of no volatility (b = 0) the process is just a deterministic function equal to the straight line xt = x0 +at, while in the case b = 0 we add noise around that line according to the Brownian motion zt . But what about the general case of variable a(xt , t) and b(xt , t)? We can try and interpret our integral equation as follows: the ds-integral is an ordinary Riemann integral, and the dzs -integral perhaps could be interpreted as a Riemann-Stieltjes integral for each trajectory. Unfortunately, this is not possible since one can show that the trajectories are of locally unbounded variation1 , and therefore the stochastic dzs -integral cannot be dened in the traditional way. To make the characteristics of the trajectories of a Brownian motion even clearer, let us take a closer look at some graphs. First, consider a deterministic smooth (dierentiable) function with a quite jagged trajectory, which could be considered similar to a Brownian motion trajectory. As we zoom in, we see that the function becomes smoother and straighter, until eventually it becomes a straight line.

n The variation of a process Xt over an interval [0, T ] is dened as sup i=1 |Xti Xti1 | where : 0 = t + 0 < t1 < ... < tn = T, so that the supremum is taken over all possible partitions of [0, T ].
1

70

J Herbert UCL 2010-11

Stochastic Methods in Finance

Dierentiable functions, however strange their global behaviour, are built from straight line segments, and classic calculus is the formal acknowledgement of this. Let us now take a look at a Brownian motion. As we zoom in we cannot obtain a straight line. The process replicates itself (self-similarity property) even if rescaling (zooming in).

Unfortunately this implies that the trajectory is not dierentiable at any t, and therefore the tools of traditional calculus cannot be applied. What we need now is stochastic calculus. We are going to look here at stochastic dierentials and in particular at It os formula for the dierential of a function of a stochastic process. Underlying the subject is the careful denition of the integral with respect to Brownian motion in (9.1). We will not look at this in detail in this introductory course, but instead will concentrate on some intuition and some tools needed for manipulating stochastic dierentials.

9.2

It os formula

It os formula is a fundamental result of stochastic calculus, and gives us an explicit understanding of the behaviour of functions of stochastic processes. The formula is also known as It os lemma, and states the following: It os formula. Consider a random variable x that follows an It o process dx = a(x, t)dt + b(x, t)dz where z is a Wiener process, a(x, t) is the drift rate and [b(x, t)]2 is the variance rate (non-constant). Consider also a continuous function G = G(x, t) twice dierentiable
71 J Herbert UCL 2010-11

Stochastic Methods in Finance

in x and once dierentiable in t. Then G is itself a random process and satises the following stochastic dierential equation: G G 1 2 G 2 G dG = a(x, t) + + b (x, t) dt + b(x, t)dz 2 x t 2 x x Therefore G also follows an It o process, with drift rate of G G 1 2 G 2 a(x, t) + + b (x, t) x t 2 x2 and variance rate of 2G 2 b (x, t) 2x The rigorous proof of the formula is outside the scope of this course. However, when we are dealing with a continuous and dierentiable function of two variables x and t, in ordinary calculus we usually express the dierential as dG = G G dx + dt x t
[ ]

This reects the fact that we are making a Taylor expansion dG = G G 1 2G 1 2G 2G 2 2 dx + dt + ( dx ) + ( dt ) + dxdt + x t 2 x2 2 t2 xt

and then neglecting the terms of second or higher order. In our case, however, we have to be careful because the variable x has special properties, since it satises the SDE above. Therefore, after we make the Taylor expansion, we have to look carefully at the second order terms to see if we can actually neglect them. It turns out that whilst the terms in (dt)2 and in dxdt are indeed of higher order and hence can be dropped, the term in (dx)2 is actually of order dt and so cannot be dropped (see Appendix for more explanation of why this is the case). Hence if we go back now to our Taylor expansion, we see that we cannot neglect anymore all the terms that looked second order, since the term in (dx)2 is actually of order dt. The terms in (dt)2 and in dxdt, instead, are of higher order, and can be dropped. Therefore we are left with dG = G G 1 2G 2 dx + dt + b (x, t)dt x t 2 x2 ()

If we substitute in the denition of dx according to the SDE of the process we obtain It os formula.
Note. Sometimes It os result is reported directly as above, i.e. stating the following: It os formula (2). If x is an It o process and G(x, t) is C 2,1 , then dG is given by (*). This is absolutely equivalent to the formula we gave earlier (just substitute in dx from the SDE). 72 J Herbert UCL 2010-11

Stochastic Methods in Finance

9.3

Examples

We are going to apply now It os formula to derive the SDE followed by some specic processes. The rst two were given without proof in the last handout, and are derived here using It os formula. The other examples are more general applications. In all these examples we always start from an It o process dx = a(x, t)dt + b(x, t)dz and then specify a(x, t) and b(x, t) in order to obtain the processes of interest.

9.3.1

Derivation of the SDE of a generalised Brownian motion

Let us start from a standard Brownian motion. In this case our process x is simply z , so the SDE will obviously be dx = dz , which is an It o process with a(x, t) = 0 and b(x, t) = 1. Let us now consider a transformation of the process with a shift, a change of scale and the addition of a trend term: G(x, t) = x0 + t + x where and are constants. We know that G follows a generalised Brownian motion, since the dierences are given by G = t + x which by denition is a generalised Brownian motion, since x in this case is a standard Brownian motion. To apply It os formula we compute: G = x G = t 2G = 0 x2 It follows that dG = dt + dz is the SDE of a generalised Brownian motion with drift and volatility .

73

J Herbert UCL 2010-11

Stochastic Methods in Finance

9.3.2

Solution of the SDE of a geometric (exponential) Brownian motion

In the last handout we said that the solution of the SDE of the geometric Brownian motion was the exponential of a generalised Brownian motion with certain parameters. Now we have the show to prove it: we start from a generalised Brownian motion satisfying the following SDE 2 dx = dt + dz 2 In this case we have an It o process with a(x, t) = 2 and b(x, t) = . From the previous example we know that the solution of the SDE is given by 2 x= t + z 2 Let us consider now the exponential transformation G(x, t) = ex We see that G is only function of x, so all the derivatives in t will be zero. Compute now: G = ex = G x 2G = ex = G 2 x Then we have 2 dG = G 2
[ ( ) ( )
2

1 + G 2 dt + Gdz 2

which simplies to dG = Gdt + Gdz i.e. the SDE of the geometric Brownian motion. So we have proved that the solution of the SDE of a geometric Brownian motion is
{(

G = exp

2 t + z 2

i.e. the exponential of a generalised Brownian motion.

74

J Herbert UCL 2010-11

Stochastic Methods in Finance

9.3.3

Logarithm of stock prices

In the last handout we established that an appropriate model for stock prices is the geometric Brownian motion. From the previous example, then, we can write
{(

S = S0 exp

2 t + z 2

and therefore we have that the logarithm of stock prices follows a generalised (not geometric!) Brownian motion. In fact: 2 Y = log S = Y0 + t + z 2 where Y0 = log S0 , i.e. a generalised Brownian motion with drift rate variance rate 2 . This result can also be obtained starting from the SDE for S dS = Sdt + Sdz and applying It os formula to the function G(S, t) = log S .
2 2

and

9.3.4

Generic transformation of stock prices

We can derive a general formula valid for any function of stock prices. Consider the usual geometric Brownian motion for S as above. We have a(S, t) = S b(S, t) = S Applying It os formula, we obtain that a generic transformation of S and time, G = G(S, t), follows a process with SDE G G G 1 2 G 2 2 dG = S + + S dt + Sdz 2 S t 2 S S
[ ]

9.3.5

Forward price

A direct application of the previous formula can be the determination of the SDE satised by the forward price. Consider a forward contract on a non-dividend paying stock. The forward price is F = SerT , where T is the time to maturity. We can express this as a function of current time t by setting T = tm t, where tm is the maturity date. We obtain F (S, t) = Ser(tm t)
75 J Herbert UCL 2010-11

Stochastic Methods in Finance

our function of interest. The process of stock prices is as usual a geometric Brownian motion. We have F S F t 2F S 2
[

= er(tm t) = rSer(tm t) = 0

From the formula derived in the previous example we have dF = er(tm t) S rSer(tm t) dt + er(tm t) Sdz Substituting F = Ser(tm t) this becomes dF = ( r)F dt + F dz i.e. F is a geometric Brownian motion too, with expected growth rate r and volatility . The growth rate of F is the excess return of S over the risk-free rate.
]

9.4

Appendix: second order eects in the limit

Here we look at the limit of dx2 . Remember that x follows an It o process, and hence small changes in x are given by x = a(x, t)t + b(x, t) t where we used the Brownian motion property z = t. Dropping the arguments of functions a and b, the dierences squared are then given by (x)2 = a2 (t)2 + b2 t2 + 2ab(t)3/2 The rst and last terms are of order higher than t, but the central term has the same order as t; we also know that 2 2 1 and therefore E [t2 ] = t Var[t2 ] = 2(t)2 As t 0, the variance of t2 will tend to zero faster than its expectation, so we can treat the term as non-stochastic and equal to its expected value t. It follows that (x)2 will tend to b2 dt as t tends to zero, i.e. (dx)2 b2 dt
76 J Herbert UCL 2010-11

Stochastic Methods in Finance

9.5

Further reading

Wilmott (ch. 7) and Hull (Section 10A in 4th edition) provide a heuristic discussion of Itos lemma similar to the one presented in these lectures, while Neftci (ch. 9 & 10) and Bjork (ch. 3) go into slightly more detail. See for example Stochastic dierential equations: An introduction with applications by Bernt Oksendal (Springer) for a more advanced treatment of the It o integral and of stochastic dierential equations.

77

J Herbert UCL 2010-11