@correlator created a chart that belongs here:
It’s the distribution of the standard deviation of TC
As it seems, lower MMcorr comes with higher TC AND with higher volatility
2 Likes
I’m not really surprised by the results, since a 100% correlation with the meta model also means most of the stake is already on that kind of prediction, which results in a low TC.
I do have one issue though. Your post title mentions “optimizing” for TC, but to really optimize for TC by trying to not be correlated with the meta model, you need the actual meta model predictions. It would be really helpful if the train/test parquet files included an additional column with these predictions.
Without that information, it feels like we’re just “cooking” - trying out anything and going with what seems to work, rather than truly optimizing.
Do you know if there’s a way to access the meta model predictions or if they’ll be included in future train/test parquet files?
They are included with the new dataset. You can download them right now → Numerai
see this: Super Massive Data: Sunshine