Regression Targets and Error Filtering for Trading Models

Summary

The article contrasts directional classification with regression for trading forecasts. A classifier predicts direction but not the size of a move, so hit rate alone can conceal the effect of gains and losses. The proposed regression targets measure future price change, either at a randomly selected horizon within a window or as the difference between the current close and the average of future closes in that window. The author notes that a single sampled future price can miss meaningful intervening moves.

For filtering, the workflow trains several CatBoost regressors on random data splits, averages their absolute prediction errors as meta-labels, and trains a final random forest only on observations below an error tolerance. The article describes training repeated models and a EURUSD hourly example over a stated historical span, then exporting models to a trading bot. It presents this as one possible design, not a validated general result; the labeling choices, thresholds, validation approach, and instrument-specific behavior require further testing.

Key ideas

• Regression targets encode the magnitude of future price changes rather than direction alone.
• Averaging future prices over a window accounts for more than one endpoint observation.
• An ensemble of regressors supplies an average absolute-error filter for training examples.
• The final model is trained on examples whose meta-label errors fall below a tolerance.
• The workflow is a proposed approach whose usefulness depends on validation and target choices.

Tags

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