How Filtering Can Break Time-Series Rolling Windows in Stock Features

Summary

The discussion examines a sequence-window feature pipeline for stock data. If rows are filtered before a rolling window is formed, dates can disappear; a nominal three-row window may then combine observations that are not three consecutive days. The answer confirms that the described module does not check time continuity and that filtering can leave gaps in an individual stock’s series, so the resulting window may not represent the intended calendar or trading-day span.

It also notes that the module flattens each window into a vector of window length times feature count, rather than retaining a separate time-by-feature shape. Whether a one-dimensional convolutional model learns temporal structure from that representation depends on the model and input interpretation. The exchange identifies risks in the pipeline but does not present a definitive repair, empirical model comparison, or evidence that a particular correction improves predictions. Researchers should therefore verify per-asset date continuity and understand the tensor shape before interpreting rolling-window features.

Key ideas

• Filtering observations before window generation can create gaps in a stock’s time series.
• The described rolling-window module does not test whether dates in a window are continuous.
• The module flattens each window into a vector rather than preserving separate time and feature axes.
• The discussion raises uncertainty about whether a one-dimensional convolution learns temporal structure from the flattened input.

Tags

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