Diagnosing Backtest and Simulated-Trade Signal Mismatches

Summary

This community post describes a discrepancy between a CNN-based stock strategy's backtest and simulated-trading signals. With a rolling window of five, both reportedly signal the same stock on January 4. With a window of ten, the simulation notification names one stock while the program log records an order for another. The author says attempts based on community suggestions have not resolved the issue and asks for an explanation.

The post is a troubleshooting question, not a solution: it does not identify the cause or provide code, configuration details, or a verified diagnosis. It nevertheless highlights a useful validation problem. Changes in rolling-window length can expose differences in data timing, model inputs, signal calculation, or order handling between backtest and simulation. These are possible avenues to investigate, not conclusions established by the post. Reproducing the mismatch with aligned timestamps, inputs, and logged model outputs would be needed to isolate the source.

Key ideas

• The reported stock selection matches with a rolling window of five but differs with a window of ten.
• The simulation notification and program log identify different stocks for the same date.
• The post offers no diagnosis, code, or confirmed resolution.
• Comparing timestamps, model inputs, signal outputs, and order records could help investigate such discrepancies.

Tags

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