Rolling StockRanker and XGBoost Training Windows for Equity Forecasts

Summary

This user post describes rolling model training for stock forecasts using two approaches. The StockRanker example updates the model annually, fitting on one year of data and predicting over the following year. The XGBoost example instead constructs monthly configurations, each using the prior six months as its training window and the next month as the prediction period. The post presents these as implementation examples rather than a complete model specification.

The author reports that StockRanker results seemed reasonable, while XGBoost results were very poor and unexplained despite several changes. No predictions, evaluation metrics, feature definitions, target construction, or model settings are supplied, so the source cannot establish why the methods differ or whether either setup performs well out of sample. It is best read as a comparison of rolling-window schedules and a troubleshooting question. The excerpt does not provide a resolution, and its configurations alone are insufficient to reproduce or validate a trading strategy.

Key ideas

• The StockRanker example retrains annually using the preceding year of data.
• The XGBoost example refreshes monthly, using a six-month training window and a one-month prediction window.
• The author reports markedly different observed results but provides no diagnostics to explain the gap.
• Feature definitions, target construction, model parameters, and evaluation metrics are absent from the post.
• Rolling date configurations specify a schedule, but do not by themselves validate a forecasting strategy.

Tags

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