Effective Memory Cross Validation for Financial Time Series

Summary

The article examines whether including all available history improves a time-series model, arguing that older financial observations may lose relevance as markets change. Its effective-memory procedure holds out a fixed later test period and repeatedly trains on progressively larger portions of the earlier data, each time adding older observations while retaining the most recent training segment. It compares out-of-sample root mean squared error across those training windows.

In the reported EURUSD experiment, using roughly eight years of data split between training and testing, the lowest error came from 80% of the training partition rather than the full partition; the test period was roughly four years. This suggests that older observations can sometimes impair forecasts and that less training history can reduce computation. The result is specific to the dataset, model setup, target, and fixed test period. The procedure is not classical k-fold cross-validation, and a single reported experiment does not establish that trimming history will help other markets or periods.

Key ideas

• Financial time series may have an effective memory because older market relationships can lose relevance.
• The procedure compares nested training windows that add progressively older data against one fixed later test set.
• Out-of-sample RMSE is used to select the training history length.
• In the reported EURUSD experiment, the minimum error occurred at 80% of the training data.
• The finding is dataset-specific and should not be treated as a universal rule to discard older data.

Tags

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