Maximum Likelihood Estimation of Brownian Motion Drift

3.6k Views Asked by At

I'm looking at times series of stock movements over 10 minute windows, and am trying to measure the "trend" of these movements.

Method A is to simply calculate $\Delta P$, the difference between the final value and the initial value.

Method B is to use MLE to calculate the maximum likelihood value of the drift coefficient using the Brownian Motion model $dP(t)=\mu P(t) dt + \sigma P(t) dB(t)$.

My intuition is that, if I am able to calculate the estimate $\hat{\mu}$, it will be a better measure of the trend since the Brownian motion is able to take into account volatility. However, another part of me thinks that $\hat{\mu}$ may actually just be a linear combo of $\Delta P$ (i.e. $\hat{\mu}$ is completely determined by the first and last values of the time series).

Any thoughts or semi-rigorous explanations would greatly appreciated! And let me know if anything else needs clarifying.

1

There are 1 best solutions below

2
On

If $\mu$ and $\sigma$ are constant, $dP = \mu\; dt + \sigma\; dW(t)$ is equivalent to $P(t) = P(0) + \mu t + \sigma W(t)$ where $W$ is Brownian motion with $W(0)=0$. Then $\Delta P = P(T) - P(0) = \mu T + \sigma W(T)$ has a normal distribution with mean $\mu T$ and standard deviation $\sigma \sqrt{T}$. The density at a given $\Delta P$ is maximized with respect to $\mu$ when $\mu = \Delta P/T$.

Now I think what you're really doing is looking at not just the initial and final values, but some in between as well, say $P(t_j) = P_j$, $j= 0 .. N$, where $0 = t_0 < \ldots < t_N = T$. Let $\Delta_j t = t_j - t_{j-1}$ and $\Delta_j P = P(t_j) - P(t_{j-1})$. By the fact that Brownian motion has independent increments, the joint distribution of $\Delta_1 P, \ldots, \Delta_N P$ is independent normal with means $\mu \Delta_1 t, \ldots, \mu \Delta_N t$ and variances $\sigma^2 \Delta_1 t, \ldots \sigma^2 \Delta_N t$.

The density of this is proportional to $$\exp\left( - \sum_j \dfrac{(\Delta_j P - \mu \Delta_j t)^2}{2\sigma^2 \Delta_j t}\right) = \exp\left((const) + \frac{\mu}{\sigma^2} \sum_j \Delta_j P - \frac{\mu^2}{2\sigma^2} \sum_{j} \Delta_j t \right) = (const) \exp((2\mu \Delta P - \mu^2 T)/\sigma^2)$$ which is maximized when $\mu = \Delta P/T$.

Thus the two approaches have the same result.