Choosing Training Windows for AI Equity Strategies
Summary
This article examines how the length of an AI model’s training window can affect an equity strategy. It compares manually rolled long windows spanning several years with shorter windows ranging from a month to a few years, and recommends evaluating each against a fixed out-of-sample period. In the reported experiments, returns and Sharpe peaked for some intermediate windows, while adding more history did not consistently improve model quality. The author also discusses changing factor weights and NDCG scores as training periods change.
The article warns that long windows can include noisy or stale patterns and encourage overfitting, while short windows can produce unstable, underfit results whose apparent gains fade. It suggests reading ranking metrics alongside training and forward returns, and considering market regime, stock universe, risk controls, and factor fit. These observations come from a particular platform and strategy setup; the suggested metric thresholds and date ranges are not established as universal rules. Training-window choice should be validated for each model and market.
Key ideas
- Compare candidate training windows on a consistent forward validation period.
- Longer histories can stabilize evaluation but may add noise and increase overfitting.
- Short windows may show attractive returns while remaining underfit and less reliable out of sample.
- Interpret ranking metrics together with forward performance, since either can give a misleading picture alone.
- Training-window suitability depends on the model, its data, initialization, and market regime.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.