Cross-Validation Filtering and Meta-Labeling with CatBoost

Summary

The article proposes using cross-validation predictions to identify training examples that a classifier repeatedly mislabels. A primary CatBoost model is trained on examples judged more predictable, while a second classifier learns across the wider dataset to decide whether the primary model should trade. This separates directional prediction from trade selection and is intended to reduce the influence of noisy or unreliable labels.

The worked example concerns EUR/USD classification, with a specified historical training window, random trade durations, and a price markup intended to account for trading costs. The author describes comparing custom backtest results with MetaTrader testing and suggests further tests on real ticks and different stop, target, and position-size settings. However, the method is presented as an experimental, trial-and-error approach, and the text supplies no quantified performance evidence. Its shuffled train/test splits and use of sequential market data also warrant careful leakage and time-series validation checks before interpreting results as causal or out of sample.

Key ideas

• Cross-validation predictions are used to flag samples that the initial classifier often misclassifies.
• A directional model is trained on retained examples, while a meta-model learns whether to permit a trade.
• The example uses CatBoost for EUR/USD classification and includes a markup parameter for estimated trading costs.
• The author recommends comparing backtests and testing alternative execution and risk settings.
• The method is exploratory and the article provides no quantified evidence that it improves trading performance.

Tags

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