Hi everyone,

I am trying to construct a metric to classify factors, maybe some of you have good idea. I would like a "simple" metric, by simple I mean not a model comparison using bayesian method (which I could "copy-past" from what we do in cosmology for model classification). I want to avoid this because I do not want to have to run to heavy computation and I do not need a "very" accurate classifier. The goal is just to be able to classify 100-1000 factors quickly and roughly.

One metric is the is the IC from alphalens, but this is not a good metric. What it does test is how the relation between the factor and the return can be explained by a monotonic function (see Spearman's rank correlation coefficient). This does not tell us about the shape of the monotonic function at all (as it is nonparametric). The issue of this test is simple. Let say that we have 2 factors, both with a IC very close to 1. But one best fit function is an arctan and the other is shifted arctan. It is obvious that the arctan one will be a better factor than the shifted arctan (one can think about the fact that the slope around 0 return is higher in the arctan than in the shifted one, which mean that the factor will be more accurate to classify positive and negative returns). But the Spearman's rank correlation coefficient will not be able to capture this fact.

I think that doing some stats on the top/button quantiles would be much better classifier than the IC from Alphalens.
Would be great if one have some idea or some link to papers (ideally arxiv) which discuss this problem.