How Training Window Length Affects Quantitative Model Results

Summary

This discussion examines whether the length of a model’s training window changes an AI strategy’s results. It describes manually rolling training for a visual template strategy, comparing longer histories of five to ten years with shorter windows ranging from one month to two years. The proposed assessment holds out a specific backtest period as validation and observes changes in factor feature weights and NDCG scores.

The reported finding is that changing the amount of training data altered weights and NDCG only modestly, and NDCG did not exceed 0.6 in the runs described. The author interprets this as evidence that more samples did not reveal a strong pattern and that noise may limit predictive signal. These are observations from one strategy setup, without details on markets, data construction, validation design, or statistical uncertainty. They do not establish a generally optimal training window or prove that the same outcome applies to other models.

Key ideas

• The discussion compares training windows from one month to ten years using manual rolling training.\nIt proposes evaluating window choices on a fixed validation period.\nFeature weights and NDCG changed with training length, but changes were described as modest.\nThe reported NDCG remained below 0.6, which the author attributes to noisy data and weak learned patterns.\nThe limited setup does not identify a universally best training window.

Tags

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