This test will be conducted by using instruments with very long history as training data. Mostly well known indices and US large cap stocks. All 40 instruments got EOD data from 1980 or longer back in time with S&P500 and Oil dating back to the early seventies.
In the first round I´m using all my 535 features and 10 of my proprietary classification calculations for the training and testing. My system development strategy is to make about 10K test and then cut the number of features in half working my way down to about 32 features that is reoccurring in all of the 16 three year testing periods from 1999-2017.
After the first round the results are quite different to the ones from the In-Sample test I did last time. Now the years during the subprime crisis are very hard to get a good risk adjusted return from. Also notably is that none of the features related to point in time seems to be in the top 20 most recurrent features in the different market regimes. It seems that the target calculation based on the shortest period of time is most successful this time.
Hire a freelance quant, contact me at: mikael[at]furesjo.se