Stochastic Processes – Math
15 September 2021In probability theory, a stochastic process (or random process), is the counterpart to a deterministic process. That is, instead of dealing with only one possible reality of how the process might evolve over time, in a stochastic process there is some uncertainty about the future evolution described by some probability distributions.
This means that even if the initial condition is known, there are many scenarios the process might go to. However, some paths can be more probables than others.
Brownian Motions
A Brownian Motion $B_{t}$, also referred to as Wiener process is a real valued stochastic process characterised by the following properties:
- $B_{0}=0$;
- $B_t$ has continuous paths, i.e., $B_{t}$ is continuous in $t$;
- $B_t$ has independent increments, i.e., for every $t>0$, the future increments $B_{t+u}-B_{t}$ are independent of past values $B_{s}$, for all $s< t$ and $u\geq 0$;
- $B_t$ has Gaussian increments, i.e., $B_{t+u}-B_{t}$ is normally distributed with mean $0$ and variance $u$, i.e., $B_{t+u}-B_t \sim{\mathcal {N}}(0,u)$. Have a quick look here to recall what does it mean.
A similar notational convention is used in stochastic calculus. When we refer to $dB_t$ as a Brownian motion, we mean that it has the properties for $B_t$ given above in the limit as $t\to 0$.
For a stochastic process, the mean change per unit time is known as the drift rate and the variance change per unit time is known as the variance rate. That is, the Brownian Motion $B_t$ introduced beforehand has a drift rate of 0 and a variance rate of 1.
Generalized Brownian Motions
In general a drift rate equal to zero means that the expected value of $B_t$ at any future time is equal to its current value. On the ther hand, a variance rate equal to $1$ means that the variance of the change in $B_t$ in a time interval of length $\Delta t$ equals $\Delta t$.
We can define a Generalized Brownian Motion for a variable $x$ in terms of $dB_t$ by using the following Stochastic Differential Equation (SDE): $$ \label{eq:gen:brown:mot} dx = a\,dt + b\, dB_t, $$ where $a$ and $b$ are two constants. Here, $a$ denotes the drift rate of the motion whereas $b^2$ is the corresponding variance rate.
In fact, given a normal random variable $\xi\sim\mathcal{N}(0,1)$, recall that since $\Delta B_t = \xi\sqrt{\Delta t}$, in the limit $\Delta t \to 0$, we can also write, $$ dx = a\,dt + b\, \xi \sqrt{dt}, $$ and then: \begin{align*} \mathbb{E}(dx) &= \mathbb{E}(a\,dt) + \mathbb{E}(b\, \xi \sqrt{dt}) = a\, dt + b\sqrt{dt}\underbrace{\mathbb{E}[\xi]}_{=0} = a\, dt, \\ \text{Var}(dx) &= \text{Var}(b\, \xi \sqrt{dt}) = b^2\, dt \underbrace{ \text{Var}(\xi) }_{=1} = b^2 \, dt. \end{align*} While the drift rate is used to model the trend of the stochastic process, the variance rate models unpredictable events occurring during this motion, i.e., how much the process oscillates around the trend.
Geometric Brownian Motions: The Process for a stock price
A stochastic process $S_t$ is said to follow a Geometric Brownian Motion if it satisfies the following SDE: $$ \label{eq:geom:brow:mot} dS_{t} = \mu S_{t}\,dt + \sigma S_t \, dB_{t}, $$ where $B_{t}$ is a Brownian Motion, $\mu$ is the (percentage) drift rate, and $\sigma^2$ is the variance rate. In this case, $\sigma$ is referred to as the (percentage) volatility.
It is typically assumed that stock prices follow a Geometric Brownian Motion. Notice that by dividing by $S_t$, we can rewrite the above SDE as $$ \label{eq:geom:brow:mot:ret} \frac{dS_{t}}{S_t} = \mu \,dt + \sigma \, dB_{t}, $$ and since $S_t$ denotes the stock price at time $t$, then $$ \label{eq:stock:rets} \frac{dS_{t}}{S_t} = \frac{S_{t+dt} - S_t}{S_t} =: r_t $$ is the percentage (or arithmetic) return $r_t$ at time $t$. We talked about price returns here.
Therefore, we see that stock returns follow a Brownian Motion where $\mu$ denotes the expected rate of return of the stock and $\sigma$ is the volatility of the stock price.
If we write the Brownian Motion $B_t$ by using $\xi\sqrt{dt}$, where $\xi\sim\mathcal{N}(0,1)$ is a normal random variable, $$ \label{eq:geom:brow:mot:ret:2} \frac{dS_{t}}{S_t} = \mu \,dt + \sigma \, \xi\sqrt{dt}, $$ then we see that, \begin{align} \mathbb{E}\left(\frac{dS_{t}}{S_t}\right) = \mathbb{E}\left[\mu \,dt\right] + \mathbb{E}\left(\sigma \, \xi\sqrt{dt}\right) = \mu \,dt + \sigma\sqrt{dt}\underbrace{\mathbb{E}(\xi)}_{=0} = \mu\,dt, \label{eq:mean:rets}\\ \text{Var}\left(\frac{dS_{t}}{S_t}\right) = \text{Var}\left(\mu \,dt + \sigma \, \xi\sqrt{dt}\right) = \sigma^2\,dt\underbrace{\text{Var}(\xi)}_{=1} = \sigma^2\,dt, \label{eq:var:rets} \end{align} that is, the percentage stock price returns $r_t=dS_{t}/S_t$ are normally distributed with mean $\mu\,dt$ and variance $\sigma^2\,dt$, i.e., $r_t \sim \mathcal{N}(\mu\,dt, \sigma^2\,dt)$.Itô's Lemma
Let us speak about a classic result known as Itô's Lemma, and see how it can be regarded as the natural extension of other well known results such as the Taylor expansion.
Itô's Lemma is a fundamental result used in finance, in particular, for option pricing. In fact, for example, the price of a stock option, or other derivative contracts, is a function of the underlying stock's price and time (and other things).
This lemma is a fundamental result in stochastic calculus that permits to understand the behaviour of a function of stochastic variables and, in particular, permits to explicitly find analytical expressions for the price of derivative contracts.
Let us first introduce the following process. An Itô Process is a Generalized Brownian Motion in which the drift rate $a$ and the (square root of the) variance rate $b$ are functions of the value of the underlying variable $x$ and time $t$, i.e., $$ \label{eq:ito:process} dx = a(x,t)\,dt + b(x,t)\, dB_t, $$ where $dB_t$ is a Brownian motion.
Note that the result implies that the function $F$ also follows an Itô process with a drift rate equal to $$ \frac{\partial F}{\partial x}a + \frac{\partial F}{\partial t} + \frac{1}{2}\frac{\partial^2 F}{\partial x^2}b^2, $$ and a variance rate equal to $$ \left(\frac{\partial F}{\partial x}\right)^2 b^2 . $$
Taylor expansion
Before proving the Itô's Lemma, let us recall some well known result about Taylor expansions.
Consider a continuous and differentiable function $F$ of one variable, say $x$. The Taylor expansion series of $F$ around a given point $x_0$ is $$ F(x) = F(x_0) + \frac{dF}{dx} (x-x_0) + \frac{1}{2}\frac{d^2 F}{dx^2}(x-x_0)^2 + \frac{1}{6}\frac{d^3 F}{dx^3}(x-x_0)^2 + \dots, $$ where we are ignoring higher order terms. If we define $\Delta x:=x-x_0$ and $\Delta F = F(x)-F(x_0)$, we can rewrite the expansion as $$ \Delta F = \frac{dF}{dx}\Delta x + \frac{1}{2}\frac{d^2 F}{dx^2}\Delta x^2 + \frac{1}{6}\frac{d^3 F}{dx^3}\Delta x^3 + \dots. $$ Likewise, for a continuous and differentiable function $F$ of two variables $x$ and $y$, the Taylor expansion is given by $$ \label{eq:taylor:2var} \Delta F = \frac{\partial F}{\partial x}\Delta x + \frac{\partial F}{\partial y}\Delta y + \frac{1}{2}\frac{\partial^2 F}{\partial x^2}\Delta x^2 + \frac{\partial^2 F}{\partial x\partial y}\Delta x\Delta y + \frac{1}{2}\frac{\partial^2 F}{\partial y^2}\Delta y^2 + \dots, $$ where, again, we are ignoring higher order terms. Note that in the limit $\Delta x, \Delta y \to 0$, the expansion simple becomes $$ \label{eq:taylor:2var:limit} dF = \frac{\partial F}{\partial x}dx + \frac{\partial F}{\partial y}dy, $$ due to the faster convergence to zero of all higher order terms.
A (not so rigorous) proof of Itô's Lemma
To prove the Itô's Lemma, let us extend the defintion of the Taylor expansions to cover functions of variables which follow an Itô process.
Suppose that a variable $x$ follows the following Itô Process $$ \label{eq:ito:proc:proff} dx = a(x,t)\,dt + b(x,t)\, dB_t. $$ The discrete version of this equation is $$ \label{eq:ito:proc:proff:disc} \Delta x = a\,\Delta t + b\, \xi\sqrt{\Delta t}, $$ where $\xi\sim\mathcal{N}(0,1)$ and where we have dropped the arguments.
Now, let $F$ be a continuous function of the Itô Process $x$ above and of time $t$. By analogy with the Taylor expansion of two variable we have seen before, the expansion series of $F$ would be $$ \label{eq:taylor:2var:ito} \Delta F = \frac{\partial F}{\partial x}\Delta x + \frac{\partial F}{\partial t}\Delta t + \frac{1}{2}\frac{\partial^2 F}{\partial x^2}\Delta x^2 + \frac{\partial^2 F}{\partial x\partial t}\Delta x\Delta t + \frac{1}{2}\frac{\partial^2 F}{\partial t^2}\Delta t^2 + \dots. $$ There is, however, an important difference between Taylor expansion and this expansion for $F$.
In the Taylor expansion, in the limiting argument to obtain $dF$ for $\Delta x, \Delta y\to 0$, the higher order terms $\Delta x^2$ and $\Delta y^2$ were ignored because they were both second-order terms and convergent to zero faster than $\Delta x$ and $\Delta y$, respectively.
Now, in our case, $\Delta t^2$ would still converge to zero faster than $\Delta t$, so there is no problem. Nonetheless, from the discrete version of the Itô Process, we have that $$ \Delta x^2 = a^2\Delta t^2 + b^2\xi^2 \Delta t + 2ab\xi \Delta t\sqrt{\Delta t}, $$ that is, $$ \label{eq:deltaxsquared} \Delta x^2 = b^2\xi^2 \Delta t \;+\; \text{terms of higher order in $\Delta t$}, $$ which shows that the term $\Delta x^2$ of our expansion for $F$ involves a component of order $\Delta t$ that cannot be ignored in the limiting argument when $\Delta x\to 0$ (that is, it would not converge to 0 too fast).
Now, recall that $\xi\sim\mathcal{N}(0,1)$, which means that $$ 1 = \text{Var}(\xi) = \mathbb{E}(\xi^2) - {\underbrace{ \mathbb{E}(\xi) }_{=0}}^2 = \mathbb{E}(\xi^2). $$ Then we have: $$ \label{eq:itoconj} \mathbb{E}(\xi^2\Delta t) = \Delta t\underbrace{ \mathbb{E}(\xi^2) }_{=1} = \Delta t. $$ Moreover, it can be showed that the variance of $\xi^2\Delta t$ equals $2\Delta t^2$. Then, it turns out that the variance of $\xi^2\Delta t$ is too small for it to have a significant stochastic component and, as a results, we can treat $\xi^2\Delta t$ as a deterministic term equal to its expected values, which is $dt$, as $\Delta t\to 0$.
Therefore, by taking the limit $\Delta t\to 0$ in equation for $\Delta x^2$, we have that $\Delta x^2$ becomes non-stochastic and then $$ \label{eq:itolimit} \lim_{\Delta t\to 0} \Delta x^2 = b^2 \lim_{\Delta t \to 0}\xi^2 \Delta t \;+\; \lim_{\Delta t \to 0}\text{terms of higher order in $\Delta t$} = b^2 dt. $$
Finally, we can now use the limiting argument $\Delta x, \Delta t \to 0$ in equation in the expansion for $F$ and obtain $$ \label{eq:ito:lemma:1} d F = \frac{\partial F}{\partial x}dx + \frac{\partial F}{\partial t}dt + \frac{1}{2}\frac{\partial^2 F}{\partial x^2}b^2 dt. $$ If we substitute the expression for the Itô' process for $dx$ we get $$ \label{eq:ito:lemma:2} \begin{split} dF = dF(x,t) &= \frac{\partial F}{\partial x}\left(a\,dt + b\,dB_t\right) + \frac{\partial F}{\partial t}dt + \frac{1}{2}\frac{\partial^2 F}{\partial x^2}b^2 dt \\ &= \left( \frac{\partial F}{\partial x}a + \frac{\partial F}{\partial t} + \frac{1}{2}\frac{\partial^2 F}{\partial x^2}b^2\right)dt + \frac{\partial F}{\partial x}b\,dB_t, \end{split} $$ which is the Itô's Lemma.