Testing Input Transformations for a Feedback-Controlled Trading Strategy

Summary

The article treats preprocessing as a tunable part of a machine-learning trading pipeline and compares transformations in a feedback-controlled strategy. Its baseline uses twelve input features without scaling; after an initial 90-day period, the controller uses observations of strategy performance to decide whether to allow trades. The underlying entry logic buys or sells when price breaks a moving-average channel.

The author tests raw inputs, z-score normalization, L1 unit scaling, and a hybrid of the two on identical historical data while holding other variables constant. The article reports that z-scoring reduced profitability and Sharpe ratio, while unit scaling improved those measures and also reduced trade count and gross loss; the hybrid performed worse than baseline on profitability and Sharpe. These are results from one described setup, not proof that unit scaling generalizes. The article offers controlled benchmarking as a practical way to assess transformations, while acknowledging that no universally best preprocessing method is established.

Key ideas

• The article frames input preprocessing as a tuning choice that can change a trading model's results.
• Its baseline strategy trades moving-average channel breaks and later uses a feedback controller to filter entries.
• The author compares raw data, z-score normalization, L1 unit scaling, and a hybrid under otherwise constant test conditions.
• In the reported test, unit scaling improved several performance measures, while the hybrid weakened results.
• The comparison is specific to this strategy and historical test; the best transformation may differ in other settings.

Tags

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