Ensemble methods improve prediction by combining the outputs of multiple models rather than relying on a single model's forecast. In quantitative signal research, ensembles are widely used to generate more stable composite alpha signals and to reduce the risk that any single model overfits in-sample.
Common ensemble techniques
- Bagging (Bootstrap AGGregatING) — trains many models on different bootstrap samples of the data and averages their predictions. Reduces variance. Random Forests are the canonical example.
- Boosting — trains models sequentially, each correcting the errors of the previous. Reduces bias. XGBoost, LightGBM, and CatBoost are widely used in finance.
- Stacking — uses a meta-model to learn the optimal combination of base models' predictions, trained on held-out data.
- IC-weighted averaging — a signal-research-specific ensemble: weight each sub-signal by its recent rolling IC, giving more influence to whichever signal has been most predictive recently.
Ensemble methods are also a form of signal aggregation: rather than deciding which single signal to use, they let the data determine the optimal blend.