Reversal Factor by Trade-Size Activity

Article BigQuant

Summary

This note describes a method for splitting a stock’s recent trading days by average transaction value, calculated as daily traded value divided by trade count. Over an even-length lookback window, it ranks days by this measure, sums returns for the higher-value half and lower-value half separately, and defines the refined reversal signal as the difference between those two return sums. The idea is that returns on high-activity days carry different information from returns on low-activity days.

For a 20-day example on A-shares, excluding special-treatment stocks and recent listings, the note reports that the high-activity component has negative rank information coefficient while the low-activity component is mildly positive over the stated historical sample. It also describes a machine-learning ranking implementation trained on earlier years and applied to later years, reporting positive relative return, but cautions that this result does not demonstrate the factor’s durable performance. The author suggests testing alternate windows, horizons, intraday data, and combinations with other factors. The factor construction and reported results are specific to the sample and implementation; further validation is needed.

Key ideas

The method ranks recent trading days by average transaction value, rather than sorting stocks by their overall turnover.
It separately sums returns from high-activity and low-activity days, then contrasts the components to form a reversal signal.
The reported historical results show stronger reversal behavior in the high-activity component and weaker momentum in the low-activity component.
A ranking model is used as a follow-on test, but the note itself questions whether its result indicates a durable effect.
Lookback window, forecast horizon, intraday inputs, and combinations with other factors are proposed as further experiments.

Refining a Stock Reversal Factor by Trade-Size Activity

Summary

Key ideas

Tags