Reflections on Regime Change and Rolling Training for AI Strategies

Summary

The author reflects on the difficulty of building a strategy that remains profitable across bull and bear markets. They identify detecting changes in market style as a promising research problem and describe the challenge of interpreting AI models whose decisions are less transparent than those of linear strategies. They also raise practical questions about which factors to use, how many to include, and whether the factors should be weakly correlated.

The post argues that models should be retrained on rolling data: training on older market features may leave a model poorly matched to newer conditions. It mentions a backtest of an AI strategy over 2024–2025 using 2022–2023 training data, but gives no performance figures or details about the model, assets, or test design. These are personal observations rather than a systematic comparison, so they illustrate concerns about regime change and model evaluation without establishing that rolling learning improves returns.

Key ideas

• Strategies may lose effectiveness as markets move between regimes.
• Detecting changes in market style is presented as a useful direction for AI research.
• The author wants AI models to be more interpretable and raises questions about factor selection and correlation.
• Training on older data may not reflect newer market conditions, motivating rolling model updates.
• The post offers personal reflections and a backtest reference but no detailed evidence or methodology.

Tags

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