I am currently working on evaluating an American Option using a Monte Carlo simulation, and I am getting answers but they vary quite a bit. The two variables that I can alter are number of simulations $N$ and number of time-steps $M$. The more time steps I use, the closer the mean of the answers are to the correct value, whereas the more simulations I use, the lower the variance is of the answer I receive.

The variance of my answer should be proportional to $\frac{1}{\sqrt{N}}$.

My question is: Is there a known method that I can use to evaluate the optimal ratio of $N:M$?

Background: To value the option, I create $N$ simulations where each one is made up of $M$ nodes, following a Brownian motion.

Thanks in advance.


You can get the relation of the optimum number of steps per run to the value $T$ of available total steps to be take ($T = NM$), under the following assumptions:

  • There is some reason to know that the finite-step-size error goes as some specific power $\frac{1}{M^p}$. For example, many simulations linearize on each step, leading to accumulated errors that go as $\frac{1}{M^2}$.

  • You may assume that that finite-step-size error is a gaussian-distributed random variate over the ensemble of runs.

  • You can estimate (perhaps by using up a small fraction of your available runs) the coeffiicent of $A_N$ of the $1/\sqrt{N}$ in your run-answer variance and the coefficient $A_M$ of $\frac{1}{M^p}$ in your finite-step-size error.

In this case, the total squared variance will be
$$\sigma^2 = A_N^2 \frac{1}{N} + A_M^2 \frac{1}{M^{2p}} = A_N^2 \frac{M}{T} + A_M^2 \frac{1}{M^{2p}} $$ Then simple calculus says the optimum number of steps in each run is at $$M = \left( \frac{2pA_M^2}{A_N^2} T \right)^{\frac{1}{2p+1}} $$

For example, if you find $A_N = A_M$ in the example with $p=2$, the optimum number of steps per run is $$ M = \sqrt[5]{4T}$$ So that if you can afford ten billion total steps, you should use about 75 million runs of runs 132 steps.

  • $\begingroup$ Thank you. I have a question about your answer, the variable $Am$ and $P$ estimate the error of the simulation whereas $An$ and $N^{-1/2}$ estimate the variance of the simulation, am I not combining two different quantities? I have done tests and found that the Error is proportional to $M^{-1/3}$, and hardly any relationship between the Error and $N$. Whereas the Variance of the answers is proportional to $N^{-1/2}$, and there is no relationship between the Variance of the answers and $M$. $\endgroup$ – Koen Rijks Jan 9 '15 at 11:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.