Selecting Regression Targets with Mutual Information for Trading Models

Summary

The article proposes using mutual information (MI) to choose which of several candidate trading targets a black-box regression model should learn. The examples compare future changes in price, moving averages, Stochastic, and RSI. MI is presented as nonparametric, unitless, and anchored at zero, making it more suitable than scale-sensitive metrics such as RMSE for comparing targets with different units. The author also argues MI is less vulnerable to reward hacking in which a model predicts near-average returns rather than responding to meaningful changes.

The workflow uses time-series cross-validation and compares targets for a statistical estimator, then builds a strategy around the target the model learns most effectively. The article reports that this revised selection approach improved backtest profit from $38.58 to $145.24 and Sharpe ratio from 0.13 to 0.4, with other listed trading settings held constant. These are results from the article's particular historical backtest, not evidence of general performance. The text is incomplete in places, so it does not provide enough detail to independently assess the full model comparison or robustness across markets and periods.

Key ideas

• Mutual information can compare regression targets whose scales and units differ.
• RMSE may favor predictions close to the training average and can obscure target informativeness.
• The example evaluates future changes in price and several technical indicators as candidate targets.
• Time-series cross-validation is used to compare models while respecting temporal order.
• The reported strategy gains are specific to the presented backtest and do not establish out-of-sample robustness.

Tags

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