Removing Overlap Bias from Rolling-Window Trading Features
Summary
Rolling estimates such as 30-day volatility share most of their underlying observations from one day to the next. A naive comparison of adjacent estimates can therefore appear highly persistent even when much of that relationship is mechanically caused by the shared data. The document illustrates this issue using GLD volatility and describes thinning the series to retain one observation per window before comparing successive estimates.
The approach reduces the sample size, so the article suggests repeating the analysis from different starting offsets to examine other sets of non-overlapping observations. Its evidence is graphical: the adjacent estimates appear strongly related, and offset samples produce somewhat different plots while showing a similar relationship. The method is a practical way to reduce dependence induced by overlapping windows, but it discards many observations and does not by itself establish that the remaining measurements are fully independent or prove a causal trading effect.
Key ideas
- Adjacent rolling-window estimates share data, which can create artificial persistence.
- For a feature using a fixed-length window, sample observations at least one window apart before comparing them.
- Apply the non-overlap filter before lagging the feature.
- Changing the sampling offset produces additional non-overlapping samples for comparison.
- Thinning reduces overlap bias but leaves fewer observations.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.