LightGBM Factor Stack · CSI 300

Most "ML in quant" tutorials are data leakage with extra steps. This template shows the validation discipline that separates a deployable model from a cherry-picked backtest.

Why this works

Tree-based models capture nonlinear interactions between factors that linear Fama–French style regressions miss — for example, the interaction between accrual quality and momentum in A-shares. Walk-forward training is the key discipline: fitting once on the whole history leaks future information. Turnover cap is what turns a paper strategy into something deployable with 15 bp round-trip costs.

Common pitfalls

1. 1Training on the full history at once. Even with cross-validation, temporal leakage destroys out-of-sample Sharpe.
2. 2Forgetting the T+1 rule on A-shares. Same-day close-to-open fills are illegal; your backtest must respect it.
3. 3Letting turnover run free. Without the cap the model wants 600% yearly turnover; nobody pays for that edge.

Try it yourself

Fork the template into your workspace. The entire configuration — code, parameters, backtest window, cost model — lands in a new private session. Tweak it, break it, and see how robust the edge actually is.