So I have been doing some reading and brain storming and have come up with a rough idea for new robust testing.

When we do a Monte Carlo analysis we have the option to randomise historical data and also the ATR of a given time period.

However this will only test the generated strategy over a given time series ie X bars.

What if we could use say X number of bars, randomise the history and volitility of those bars to generate let’s say 100 different data sets of historical data. We then generate strategies which are fitted to those historical data sets ie we find strategies that meet our acceptance criteria over those simulated data sets by taking a mean of the back tested results of a strategy over all data sets, much like in Monte Carlo simulations where we find a mean over all simulations to generate some sort of expected outcome into the future.

This way rather than generating strategies on one data set then testing over either OOS, Monte Carlo variations and multi market we are actually training our EA’s over many data sets which have a higher probability of forecasting future what the market will do in the future. Because we are taking a variance of volitilty in the present/recent data and training over many different combinations.

It is a much higher probability that say the volatility of a given market will remain within some window of variance of volatility in the coming future than say training a strategy over what volatility was like 10-9-8-7 etc years ago.

]]>