Can we get some variant of “if MMC < 0 then min(0, max(CORR, MMC)), else MMC” to get less burned if your CORR is better than your MMC?
So we had people ask about some variant of MMC scoring that doesn’t burn people that try to make a model that does well during burns (me included) - we just went through a phase where a CORR of 0.012 or so would get you a negative MMC of -0.033, because you are outcompeted by models that do well now, but would burn heavily during burn times. That seems unwanted behavior long-term for numer.ai, and is a strong disincentive to stake MMC with models that try to do well across all eras, as you burn when things are good, and you won’t make up for that during the shorter periods when things are bad and your model’s CORR still does consistently OK.
The above formula is quite nice in that there is no weird jumps in behavior as CORR or MMC goes above/below 0.0. What it does is that when MMC is above 0, your score is MMC. When MMC is below 0 and CORR is above 0, your score is 0. When MMC and CORR are below zero, your score is the max of the two. The last of those rules prevent the formula from having a weird jumping behavior the moment your CORR crosses to below 0, but you could tinker with that part and make it “When MMC and CORR are below zero, your score is CORR”.