Although we all have different models, we’re all using the same training and validation data, so it’s hard not to have correlation between models.
And you’re free to use a different model to generate your predictions each week. If you truly feel that your current model may generate negative returns for the round, you could swap it out for a different model. Which makes you a manual decision tree I guess. You can also ensemble different models.
One idea that I just had is maybe you can try correlating your model or the metamodel’s live returns with other indices, like the S&P 500, or VIX (volatility). Although the data is encrypted and we don’t know what the targets actually mean, I imagine there must be some correlation to something in the real market.
A possible hypothesis could be that when VIX is high and the market is volatile, your model with feature neutralization works better, but during other times maybe your model that has a higher max feature exposure works better. This is a hypothesis, please don’t use it in real life.