I’m trying to understand what kind of return target_ender_60 ranks, so I know what to expect from models of extreme outcomes (the 0.0 and 1.0 bins).
My understanding: each era, stocks are ranked by a residualised forward return of roughly 60 trading days, then binned 5/20/50/20/5% into 0, 0.25, 0.5, 0.75 and 1. Is that right?
The questions:
- Volatility scaling. Before ranking, is each stock’s return divided by its own volatility (a z-score or Sharpe-like return), or neutralised against a volatility or risk factor? Or is it a plain residual return, so that naturally volatile stocks land in the tails more often?
- Residualisation. Which exposures are removed: market beta, country, sector or industry, size, other style factors? Roughly how are they removed (one regression per era, or a risk model)?
- Horizon. Is “60” trading days, starting from the era date or after a short delay?
Why I ask: on raw S&P 500 prices, a stock’s trailing volatility alone puts about 30% of the most volatile tenth into the top or bottom 5% of residual returns. With Numerai’s features, my models get to about 17%. If the target is volatility-scaled, that gap is by design and nothing is being missed. If not, the features may be carrying less volatility information than raw prices do.
Even a yes or no on question 1 would help a lot. Thanks!
Plain version: I’m asking whether Numerai grades each stock’s move relative to how jumpy that stock normally is. If it does, a jumpy stock doesn’t get an advantage at landing in the extreme bins, and my models are already near the limit.