Fading the ML signal

hmmm. I’d only flip the sign if i had strong evidence the underlying relationship had structurally inverted, not just because of a drawdown.

In the AlphaNova competition, which uses walk-forward evaluation, your XGBoost model must be trained to recognize stable cross-sectional patterns. Reactively flipping signs will only damage your long-term Sharpe ratio metrics on hidden test data.