# FxPaul

Math in finance or vice versa

with one comment

## Introduction

In mathematics, Ornstein-Uhlenbeck process satisfies the following stochastic differential equation:
$dS = \theta(\mu - S) dt + \sigma dW_t$
where $\theta$ – mean reversion parameter, $\mu$ – mean and $\sigma$ – volatility.

In finance, it is used to model interest rates, currency exchange rates and commodity prices. Although it is usually modified to incorporate non-negativity of prices.

## Ordinary Least-Squares Approach to calibration

The simplest approach to the calibration problem is to convert SDE to finite difference equation (as it is usually used in Monte Carlo simulation) and to rearrange parts to Ordinary Least Squares equation.

The simplest updating formula for Ornstein-Uhlenbeck process is:
$S_t - S_{t-1} = \theta (\mu - S_{t-1}) \Delta t + \sigma \sqrt{\Delta t} W_{t-1}$
By rearranging we obtain:
$S_t - S_{t-1} = \theta\mu\Delta t - \theta S_{t-1}\Delta t + \sigma \sqrt{\Delta t} W_{t-1}$
Comparing with simple regression formula:
$y = a + bx + \epsilon$
we can equate as follows:
$y = S_t - S_{t-1}$
$x = S_{t-1}$
$a = \theta\mu\Delta t$
$b = -\theta\Delta t$
$\epsilon = \sigma \sqrt{\Delta t} W_{t-1}$
and immediately obtain the following:
$\theta = -\frac{b}{\Delta t}$
$\mu = -\frac{a}{b}$

As $W_t$ is drawn from normal distribution, its expectation equals zero and one should use variance to obtain $\sigma$:
$\sigma_{\epsilon} = \sigma \sqrt{\Delta t} \sigma_W$
where $\sigma_{W} = 1$ as it has been already normalized by $\sqrt{\Delta t}$. Finally, we can obtain:
$\sigma = \frac{\sigma_{\epsilon}}{\sqrt{\Delta t}}$

So, regression of $S_{t-1}$ against $S_t - S_{t-1}$ gives estimation of process parameters.

## The modified process

Let’s consider the process with slight modification and apply the same approach to the modified process:
$dS = \theta(\mu - S) dt + \sigma S dW_t$
Then the naive updating formula is
$S_t - S_{t-1} = \theta (\mu - S_{t-1}) \Delta t + \sigma S_{t-1} \sqrt{\Delta t} W_{t-1}$
Then dividing by $S_{t-1}$:
$\frac{S_t}{S_{t-1}} - 1 = \theta\mu\Delta t\frac{1}{S_{t-1}} - \theta \Delta t + \sigma \sqrt{\Delta t} W_{t-1}$
Given simple regression formula:
$y = a + bx + \epsilon$
we can equate as follows:
$y = \frac{S_t}{S_{t-1}} -1$
$x = \frac{1}{S_{t-1}}$
$a = - \theta\Delta t$
$b = \theta\mu\Delta t$
$\epsilon = \sigma \sqrt{\Delta t} W_{t-1}$
and immediately obtain the following:
$\theta = - \frac{a}{\Delta t}$
$\mu = - \frac{b}{a}$
Applying the same logic as in previous section, finally we get:
$\sigma = \frac{\sigma_{\epsilon}}{\sqrt{\Delta t}}$

Therefore, regression of $\frac{S_t}{S_{t-1}} - 1$ against $\frac{1}{S_{t-1}}$ yields estimation of modified process parameters.

## Open questions

1. Bias of the estimators. For the original process this approach usually gives quite precise estimation of mean and volatility but fails to provide mean reversion parameter
2. Closed-form solution of the modified SDE. It could be used to improve the updating formula
3. Statistical hypothesis testing if the sample drawn from the process. This is quite crucial point as it helps to identify model regime shift in trading.