New backtesting with new more rigorous methodology is lookin' fucking sick.
We abandoned trying to find any kind of correlation between returns during different time periods, because there wasn't any.
So instead we select ten models at random, average their positions together, and use them for a two week period before rebuilding all the models and doing it again.
Thus, the results of our backtesting represent exactly what we would have earned on a given equity in a given year, because this methodology eliminates any hindsight whatsoever.
So far our average annual return is looking like 75%, but it's still very early. It's much more labor intensive to test using this process, since we have to build and backtest 26 sets of models for every year (and there's 864 models per set, which means ~22,000 models per year, which ALL have to have to be backtested).
Exciting stuff! Will keep you guys posted.