Back to Community
Feature request - multiple runs for testing non-deterministic algorithms

A few of the randomness-inspired algorithms of discussion lately ( or would benefit from mulitple-runs type capability. By this I mean the user would be able to select some # of backtest-runs to execute, instead of just one, and the resulting data and metrics be available as a data frame.

Running a non-deterministic algorithm once only really gives you a snapshot, what you really want to do is run it thousands of times and be able to measure the variance of variables between runs.

2 responses

Hello Tyler & all,

I agree...being able to run algorithms in batches (sequentially or in parallel) with downloadable results would be an improvement, although it might not be the most efficient approach if you are interested in thousands of iterations. The infrastructure is available, as Thomas Wiecki illustrated with a zipline example utilizing cloud computing to optimize a parameter (see It would be interesting to understand if Quantopian is scalable in this fashion, with its security constraints and business model of providing free backtesting capability.

Perhaps someone has experience applying Monte Carlo methods in the context of algorithmic trading? The online, browser-based version of Quantopian is not well-suited for this kind of approach, but I can imagine zipline being used offline.


Thanks for the feature request. It makes sense to me. I'll put it in our future-feature list.