Feature Extraction Mismatches and Unstable Trading Signals in Notebooks

Summary

This forum post reports that two BigQuant notebooks produce different numbers of extracted features even when no filters are applied. The author observes that one notebook appears to include Beijing Stock Exchange data while another does not, and questions whether missing values should be handled later in a separate filtering step. The post also connects the discrepancy to repeated simulation runs in which the same strategy generates different trading signals.

The document is a problem report rather than a diagnosis or solution: it contains no response explaining the platform behavior, no reproducible example beyond the described comparison, and no proposed checks. Still, it highlights a practical research concern. Differences in universe coverage, feature extraction, or data preparation can change a strategy's inputs and make simulations hard to reproduce. Researchers should verify that notebooks use matching data sources, dates, instruments, and preprocessing settings before attributing signal changes to the strategy itself. The reported cause remains uncertain.

Key ideas

• The author reports different feature counts from two notebooks with no explicit filters.
• One notebook reportedly includes Beijing Stock Exchange stocks while the other does not.
• The author suspects extraction differences may contribute to inconsistent signals across repeated simulations.
• The post offers no confirmed cause, technical explanation, or resolution.

Tags

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