Just moving this thread from RC to the forum…
Apparently there might be a “big bounty for anyone who can come up with a bonus scheme that is actually sybil resistant and aligns perfectly with [Numerai’s] incentives / safety requirements.”
@slyfox, any other requirements we should be aware of?
Suggestions so far:
Joakim (me): Pay bonuses to top k live CORR models (after n rounds) that stay within their Bayesian probability ‘cone’, based on priors from the model’s validation set predictions. This model would NOT be getting a bonus (flipped the chart to hopefully make it more clear).
@swish : Apply erasure to the top performance bonuses, any taker can be griefed by CoE if an attack was exposed by somebody, who would be rewarded in return for that.