Someone might think that improving CORR by outperforming the Example model is also improving MMC somehow. But here we have to distinguish two ways of improving CORR:
1.- If you move from Example model to MM, you are improving CORR because you are discovering what other users discovered before and you are applying it as well. You are doing something different from the Example but nothing different from the community. Your MMC will be diluted.
2.- If you do something really original that nobody in the community has done before, you will be increasing MMC a lot because your model will contribute to the improvement of the MM performance.
I think the first way of improving CORR is much easier than the second, because if someone has discovered that improvement before, more users are more likely to do it as well. However, only the second way, the difficult and original one, increases MMC significantly.
Now, the incentive is for users to improve their models by bringing them closer to the current MM. Increase CORR but not MMC. If Numerai wants to increase the MM performance, they should increase MMC incentives while reducing CORR incentives to the point where both incentives really work. In line with this reasoning, I posted this proposal a few weeks ago:
Finally, some data that supports what I say:
1.- If Example model ranking is 29th, there are 28 model that outperform the Example in CORR terms.
2.- But there are only 2 or 3 models with a significat MMC, according to Master_key article: “how to be paid for being different …”.
Therefore, most of the good models are becoming closer to the MM, but very few increase their MMC. It seems as if all good models improve in the same direction by becoming clones of the MM but not improving the MM itself. This demonstrates the lack of incentives to MMC. I think the new MMC staking payout scheme starting on round 222 is not enough. In a few months we will see if the best models are just becoming clones of the MM or are really improving it.