Four Approaches to Labeling Returns for Quantitative Models

Summary

This article explains how to turn future returns into labels for supervised trading models, arguing that the label should match the instrument, trading horizon, liquidity, and costs of the intended strategy. It first describes using an n-period return as a continuous target, then outlines binary labels that mark positive and negative returns. Binary labels are simple but discard the size of the move, treating small and large gains alike.

It next presents threshold-based three-class labeling, assigning negative, neutral, or positive outcomes according to return cutoffs. Fixed cutoffs can be poorly suited to changing volatility or market regimes, so the article suggests adjusting thresholds to volatility. A rolling quantile method instead sorts returns within a moving window into buckets to adapt to recent conditions. It also recommends inspecting skewed, fat-tailed return distributions and considering transformations and outlier treatment. The article supplies no comparative model results, and its examples do not establish that any labeling scheme will improve live trading.

Key ideas

• Return labels should reflect the strategy’s horizon, instrument, liquidity, and trading costs.
• Continuous future returns preserve magnitude, while binary labels discard differences in move size.
• Threshold-based three-class labels distinguish negative, neutral, and positive outcomes but fixed thresholds may not adapt to regimes.
• Rolling quantile labels use recent return distributions to adjust category boundaries.
• Return distributions may require transformation and outlier handling before labeling.

Tags

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