"Instead of trying to identify profitable trading algos on in-sample data that validate out-of-sample and remain profitable forward, one could instead try to identify unprofitable algos in some data sample that turn profitable in a forward sample. This often works because markets have become more mean-reverting in recent years."
Article
Has anyone tried this approach?
I've toyed with it. There are weak mean reverting effects in some models, the emphasis being weak.
I'm very uncomfortable with it, due to the explosion in the likelihood of overfitting.
Make it simple suppose there are two 'states' in the market, where different models are profitable.
You could fit an average model with N parameters. Or you could fit a model to each state, so you have 2N parameters. You then need a model to determine which state you're likely to be in over the period you're going to hold your positions for. That will probably need at least one parameter.
So we've gone from N to 2N+1 parameters.
Of course if you're fitting method is robust enough, properly out of sample, blah, blah, blah.... .
The other alternative is to make the two models much simpler with say N/3 parameters each. Thus you end up with the same or slightly fewer parameters; and you've effectively gone from capturing behaviour in one way, to capturing it in a more non linear way. Once again I've never seen this work that well, as you often end up with something very similar to the original one state model.
"This often works because markets have become more mean-reverting in recent years." - obvious question, why not fit the mean reversion directly?
