Using a Neural Network to Rank Stocks for Rapid Rotation

Summary

This practitioner report describes experiments using a deep neural network to select stocks for a rapid rotation strategy. The model labels stocks by their subsequent return over a short holding interval, divides outcomes into ranked classes, and uses a broad set of equity factors. The author trains on a recent historical window, removes selected securities such as suspended and specially treated stocks, and tests alternatives with different factor counts and stock universes. The trading setup holds two stocks at a time and sells each after the next day’s close, with attention to balancing cash and position sizes.

The reported experiments are limited and mixed. The author observes that the strategies tended to rise during market declines and fall during advances, and that simulated results underperformed the broad market. Repeated runs also produced different outcomes unless the model was fixed. The article does not provide full performance statistics, and the author is uncertain whether the effect comes from model ranking, the rotation design, or other implementation choices. It presents exploratory findings rather than a validated strategy.

Key ideas

• The model converts short-horizon future returns into ranked outcome classes for stock selection.
• The experiments compare different factor sets and stock universes within a two-stock rotation strategy.
• The author recommends balancing cash and holdings to avoid uneven position weights during daily turnover.
• Repeated model runs sometimes produced different results, while fixing the model reduced that variability.
• The reported strategies lagged the market and appeared to perform better during market declines.

Tags

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