How Filtering and Normalization Affect LSTM Sequence Windows

Summary

This forum post raises a data-preparation problem in an LSTM trading strategy. The author reports that filtering rows before rolling sequence windows can remove observations that should remain part of the temporal context, causing later observations to shift into unintended positions. The example asks whether excluded close-price observations can still be used inside windows formed from neighboring observations, even when they are not retained as training or backtest samples.

It also asks whether filtering can be applied after window construction and normalization, and whether normalization can instead follow windowing. The document contains the question and a reference to the strategy, but no response, worked example, or resolution. It therefore identifies an important pipeline-ordering issue without establishing a recommended implementation. Researchers would need to define which records are excluded as samples versus retained as context, and guard against leakage when fitting normalization parameters.

Key ideas

• Filtering observations before rolling windows can alter the intended temporal sequence.
• The author distinguishes excluding a row as a sample from retaining it as context within a sequence.
• The post asks whether filtering should occur before or after windowing and normalization.
• No solution or example is provided, and normalization choices should be checked for data leakage.

Tags

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