Improving AlphaNet Stock Selection with Temporal Layers and Feature Changes

Summary

This report summary describes revisions to AlphaNet, a neural network that learns stock-selection factors from price and volume data. AlphaNet-v2 adds six ratio features, replaces pooling and fully connected layers with an LSTM to capture temporal patterns, and allocates a larger share of data to training while emphasizing more recent validation samples. AlphaNet-v3 adds feature extraction layers with different lookback windows and replaces the LSTM with a smaller GRU.

Tests cover Chinese broad-market, CSI 800, and CSI 500 universes over January 2011 through July 2020, with rebalancing every ten trading days. The summary reports improvements in RankIC, ICIR, benchmark-relative excess return, and information ratio for the model variants in the specified universes. These are historical backtest comparisons and do not establish future performance. The report also contrasts end-to-end learning, which reduces manual factor maintenance but is less interpretable, with genetic programming plus random forests, which offer different strengths and limitations.

Key ideas

• AlphaNet-v2 adds ratio features and uses an LSTM to model temporal information in market data.
• AlphaNet-v3 uses feature extraction with different lookback windows and replaces the LSTM with a GRU.
• The reported historical tests show improvements for v2 over v1 in the broad-market and CSI 800 universes.
• The CSI 500 comparison reports a modest improvement for v3 over v2.
• End-to-end neural models reduce manual factor engineering but are less interpretable than the compared genetic programming approach.

Tags

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