DeepAlpha DNN and CNN Stock Selection Across Multiple Holdings

Summary

This note compares DeepAlpha deep neural network and convolutional neural network stock selection experiments in a small-cap universe. The author describes a daily strategy that buys five stocks and sells them the next day, using 57 factors drawn from 19 daily factors over three days. The universe is limited to smaller stocks and filtered for trading activity. Expanding the number of selected stocks to 50 reduced returns in both model types, according to the author, who says performance came mainly from the highest-ranked names.

The author reports that adding sector or concept factors improved the ability to select many stocks, while warning that concept information may become stale. An observation that returns differ substantially across rank positions motivates a hypothesis: different ranks may represent distinct profitable styles, so training a diversified strategy involves balancing those patterns. The note provides no detailed performance tables, test period, transaction costs, or statistical validation, so its conclusions are exploratory and may not generalize.

Key ideas

• The experiments compare DNN and CNN models for next-day stock selection in a small-cap universe.
• Selecting more stocks was associated with lower returns in the reported tests.
• Sector or concept factors appeared to improve multi-stock selection.
• Returns differed across rank positions, suggesting that score order alone may not identify the best holdings.
• The author proposes balancing multiple profitable styles when training a multi-stock strategy.

Tags

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