Improving AlphaNet Stock Selection with Temporal Layers and Features

Summary

This research note describes two revisions to AlphaNet, a neural model that learns stock selection factors from raw price and volume data. Version two adds ratio features, replaces pooling and dense layers with an LSTM to capture temporal patterns, and gives more weight to recent validation samples. Version three expands feature extraction across different lookback windows and replaces the LSTM with a smaller GRU model.

Backtests over 2011 to mid-2020 report better rank information coefficients and neutralized portfolio results for version two over version one in the broad A-share universe and CSI 800; version three is modestly better than version two in the CSI 500 universe. The note also contrasts end-to-end AlphaNet with genetic programming plus random forests: AlphaNet reduces factor-pipeline maintenance and can be adapted across universes, horizons, and frequencies, but is less interpretable and supports fewer factor operations. These are historical backtests, so they do not establish future performance.

Key ideas

• AlphaNet version two adds ratio inputs and uses an LSTM to model feature sequences.
• Version three uses multiple feature lookback windows and a GRU to reduce model parameters.
• Reported backtests show version two improving on version one in the broad A-share and CSI 800 universes.
• Version three shows a smaller improvement over version two in the CSI 500 universe.
• End-to-end learning simplifies factor maintenance but reduces model interpretability.

Tags

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