Finding Forex Seasonal Patterns with CatBoost and Time Filters

Summary

The article describes a machine learning workflow for searching for time-based patterns in forex trading. A filter selects bars by hour or weekday, and the labeling process keeps only examples that meet that condition. The same filter restricts trade entries during testing, while exits can occur at any time. An exploratory routine trains models repeatedly across candidate time periods and compares their R² scores to identify hours or weekdays with more consistent results.

The examples use GBPUSD hourly data and report that some hour clusters produced denser, stronger scores, while other periods had more variable results. Tests across weekday clusters and longer historical windows show that a selected model can perform poorly on earlier data, which the author attributes as a possibility to changes in market structure. The article recommends checking multiple periods and retraining only after assessing stability. It gives no detailed statistical controls for multiple comparisons or broader out-of-sample validation, so the reported patterns should be treated as exploratory rather than established evidence of durable profitability.

Key ideas

• A time filter can restrict both training examples and the times when a strategy opens trades.
• Repeated model training across candidate hours or weekdays helps compare the consistency of discovered patterns.
• The GBPUSD examples show that some time clusters score more consistently than others.
• Results on older historical data can differ sharply from more recent performance.
• The article presents the workflow as exploratory and does not establish that selected patterns will persist.

Tags

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