ML Factor Ensemble

Use this skill to combine many existing alpha factors into a single out-of-sample meta-signal, using machine learning with strict leakage control. It sits downstream of the factor libraries (which produce individual factor columns) and upstream of evaluation (skill-ic-analysis) and backtest (skill-backtest).

This is supervised combination, fundamentally different from skill-factormad-debate-factor-mining (LLM debate that generates new factor formulas) and from skill-factor-evaluate (scores a single factor). Here the factors already exist; the job is to learn their weighting without look-ahead.

Why leakage control is the whole point

Naively fitting an ML model on a stacked panel and reading cross-validated scores massively overstates performance, because overlapping forward-return windows leak future information into the training folds. This skill enforces: