Following my post on Numerai-induced ticker churn, I’ve decided to look into what’s the churn for Numerai targets in the v1 data. After all, that’s what our models try to aim for.
Here are the results: in a perfect world, Numerai would want us to have 40.7% churn on average (or 28.3% at the very least). Code and stats:
Disclaimer: yes, having big ticker gaps across eras (outlined in the post above as ticker churn) would make the target churn worse since it could compare to an era that was over a week ago, but this shouldn’t distort results much realistically since you cannot trade a ticker if it regularly falls in and out from the universe. (so at the end of the day, ticker churn is basically additive)
To the team: I remember you mentioned that targets in the main tournament are adjusted for churn. Assuming you want us to have both high corr with the target and low churn, would it be possible to adjust targets for churn in the V2 signals data too?
Thanks.