2
$\begingroup$

I use a Monte Carlo simulation (say 100.000 runs) to estimate parameter in R. I have memory problems and my first thought is to run multiples times my estimation program (say 500 times) .

My question are: what is the distribution of an estimator ? how to get an estimator from multiple estimators ? (for the mean you can use a mean of means for example, does this hold in general ?). Is it possible to derive some kind of confidence interval from the runs ?

My questions are both general and specific to quantile and correlation estimation. For the quantile and correlation part thanks to assume a general distribution.

$\endgroup$
2
  • 1
    $\begingroup$ Your description is pretty vague but it seems that is is just a programming issue with efficient using of memory (i.e. not storing all the simulation output in RAM but taking only the important or aggregated values, saving things on disc rather that storing in memory, in most cases you can even stop a simulation and start again in the place you stopped) $\endgroup$
    – Tim
    Nov 13, 2015 at 15:49
  • $\begingroup$ Once I understand how to aggregate estimator for quantile and correlations I may be able to use inline calculation. Like after n simulation, a run r, and one estimator of the mean m I can use (n*m + r)/(n+1). For the moment the problem is I don't know how to aggregate quantile and correlation estimators (well in fact thinking about correlation, it look like a mean). $\endgroup$ Nov 13, 2015 at 15:57

1 Answer 1

3
$\begingroup$

Only the mean of means allows for a strict recovery of multiple simulations with limited storage capacities. However, Monte Carlo being an approximation method, approximations are always available.

The theoretical basis of Monte Carlo estimation is the Law of Large Numbers. This means that the empirical cdf $$\hat{F}_i(x)=\dfrac{1}{n}\sum_{i=1}^n \mathbb{I}_{x_i\le x}$$ is an unbiased estimator of the true cdf $F$. Unfortunately, you cannot store this empirical cdf $\hat{F}_i$ without storing the entire simulation sample of the $n$ $x_i$'s. However, if you replace $\hat{F}_i$ with an approximation based on the empirical percentiles $$\hat{\hat{F}}_i(x)=\dfrac{1}{99}\sum_{i=1}^{99} \mathbb{I}_{\hat{c}_i\le x}$$where $\hat{c}_i$ is the $i$-th empirical percentile, given by $$\hat{F}_i(\hat{c}_i)=\dfrac{i}{100}$$ you only need to store $99$ values. (If the percentile precision is not sufficient, you can move to the permile precision or beyond.) You can then repeat simulations without undue pressure on storage by averaging the $\hat{\hat{F}}_i$'s. From this average, you can derive a convergent estimator of the quantiles of interest.

As you noticed, since $\text{cor}(X,Y)$ can be consistently estimated as $$\dfrac{\frac{1}{n}\sum_{i=1}^n x_iy_i -\frac{1}{n}\sum_{i=1}^n x_i\,\frac{1}{n}\sum_{i=1}^n y_i}{\left\{\frac{1}{n}\sum_{i=1}^n x_i^2-\frac{1}{n^2}(\sum_{i=1}^n x_i)^2\right\}^{1/2}\left\{\frac{1}{n}\sum_{i=1}^n y_i^2-\frac{1}{n^2}(\sum_{i=1}^n y_i)^2\right\}^{1/2}}$$you can easily update the five quantities involved in this expression $$\sum_{i=1}^n x_iy_i,\ \sum_{i=1}^n x_i,\ \sum_{i=1}^n y_i,\ \sum_{i=1}^n x_i^2,\ \sum_{i=1}^n y_i^2$$ without increase storage pressure.

$\endgroup$
1
  • $\begingroup$ Thanks that was the answer I was looking for. After working a bit more on my problem, I have concluded that I have other bottleneck. Basically i am simulating a "big" sum of variables that have a t-copula dependence structure and differents compound poisson marginals, so the main bottleneck is not the quantile of my final distribution but the whole cdf estimation of my numerous marginals... If I understand correctly I can deal with my memory problems with averaging my estimators. $\endgroup$ Nov 14, 2015 at 11:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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