Meta-Labeling Bollinger Band Trades with Regime Features and Bet Sizing

Summary

This article presents a two-stage machine learning workflow for filtering Bollinger Band mean-reversion signals. A primary strategy supplies direction when price reaches an outer band; a secondary classifier estimates whether to take that trade. Its predicted probability can also inform position size, so less confident signals receive smaller allocations. The approach targets a known weakness of band strategies: touches during strong or expanding trends may signal continuation rather than reversal.

Feature engineering emphasizes band position and normalized bandwidth, together with bandwidth changes, trend strength, momentum, volatility, prior signals, and spread. Features are lagged to avoid using information from the signal bar, and the bandwidth regime threshold is fit using training data to limit leakage. The article describes triple-barrier labels, rolling performance features, and an ONNX deployment architecture. It illustrates regime effects and reports implementation details, but the supplied text does not include the full results section or quantitative performance evidence. Its conclusions therefore cannot establish profitability, and model performance depends on sound temporal validation and consistent feature handling in deployment.

Key ideas

• The primary Bollinger strategy supplies trade direction, while a secondary classifier decides whether to act.
• Bandwidth and its changes help characterize whether a band touch occurs in a ranging or expanding market.
• Lagged features and training-only regime thresholds help reduce look-ahead leakage.
• Secondary model probabilities can be mapped to position sizes.
• The provided text describes the architecture but omits detailed quantitative results.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.