MMC provides a strong incentive for users to create unique models that are orthogonal to the metamodel. More diverse predictions create a better ensemble which is good for the fund. Instead of relying on incentive structures alone to increase model diversity, Numerai should publish more information that users can use to make models more orthagonal to the metamodel.
I propose that Numerai publishes the feature exposures of the metamodel after every round. If a user knows what features the metamodel is likely to be exposed to in the next round, MMC would incentivize them to make a model that doesn’t have those same exposures. If enough staked models do this, the meta model would then be less exposed than it was in the previous round. Over time the feature exposure of the metamodel itself would fall considerably, and be less vulnerable to severe burns cries in round 260.
There is also a steep learning curve for this tournament and we should make it as easy as possible for new users start making good models. I think giving everyone more information that can easily be used to make better models can help with this.
I tried to think of ways this could be gamed to harm the tournament but I can’t think of any.
more on feature exposure: jrb’s post on feature exposure
- Good idea
- I see a problem