Why Retraining Can Change Stock Selections and Backtest Results

Summary

The document addresses a stock-selection model whose chosen stocks and reported backtest returns change when the backtest end date is moved forward by about twenty days. The explanation is that the training data differs across the runs. A model trained on a different sample can produce different fitted parameters and predictions, which in turn changes its selections and measured performance.

The example reports a large difference in annualized return between two nearby end dates, illustrating sensitivity to the evaluation window. It does not show the training procedure, feature set, date handling, or whether the backtest avoids look-ahead bias. Nor does it establish whether the changed result comes only from retraining or also from changed market observations. The note highlights model and backtest instability, but offers no diagnostic workflow or remedy for isolating these effects.

Key ideas

• Changing the training sample can change a fitted stock-selection model.
• Different fitted models may produce different predictions and selected stocks.
• The example describes a sharp annualized-return difference after moving the backtest end date forward.
• The note does not provide enough detail to distinguish retraining effects from other evaluation changes.

Tags

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