Testing Minute-Level Stock Features with Long-Short and Layered Returns

Summary

This article describes a way to assess whether a feature built from recent stock price and volume data helps predict returns over the next five minutes. Using minute data for the 50 constituents of the SSE 50, it aligns feature values with future-return labels, then calculates cross-sectional correlations at each minute. It also sorts stocks by feature value, compares a long position in the lower half with a short position in the upper half, and groups values into ten layers to examine average future returns.

The reported example shows correlations concentrated below zero, with an average near -0.2, and says the long-short series rises steadily during the observed day. The lowest and highest feature layers also appear distinct from the middle groups. These are exploratory results from one day and one index sample. The author cautions that the construction is for evaluating predictive separation, not a deployable strategy: it assumes shorting is possible and omits transaction costs. Care is also needed to avoid look-ahead in feature construction.

Key ideas

• The example uses recent minute-level price and volume data to predict five-minute stock returns.
• It evaluates the feature with minute-by-minute cross-sectional correlations, a long-short split, and ten return layers.
• The reported correlations are mostly negative, averaging about -0.2 in the example.
• The article reports a rising long-short return series and stronger separation at the extreme feature layers.
• The analysis is a one-day predictive study that omits transaction costs and assumes shorting is available.

Tags

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