Following is an update on my results from testing each algo-asset combination using the random strategy approach detailed above. I think the results both provide insight into the strategy development process and provide a useful tool for helping to find the combinations that are most likely to be profitable in real trading.

Of the 36 asset-algo combinations that were profitable in the WFO test, 10 had a backtested profit factor better than 95% of random strategies. An additional 5 combinations were better than 90% of random strategies. One of the worst performers was better than only 17% of random strategies, but had a backtested profit factor of 1.1!

The conclusions that I draw from this analysis are:

1. Clearly, data snooping bias crept into the development process and many asset-algo combinations were curve fit to the underlying trend in the market, rather than exploiting a tradable inefficiency.

2. This method was akin to taking a razor to the strategy and cutting away anything that is unlikely to do better than a random strategy. In some ways, I think that it provides a check and balance to the almost inevitable data snooping biases that creep into the development process. It therefore has merit as one of the final checks a strategy must pass before progressing to live trading.

3. 10 asset-algo combinations are better than random at the 95% confidence interval and 15 are better than random at the 90% confidence interval, estimated using an empirical approach.

If anyone has any feedback, comments or suggestions for improvement, I'd love to hear them. Thanks for reading and I hope this is of some value to the forum.