Improving AlphaNet with Sequential Layers and Expanded Price-Volume Features

Summary

This report summary describes revisions to AlphaNet, a neural network that learns stock-selection factors directly from raw price and volume data. AlphaNet v2 adds six ratio features, replaces pooling and dense layers with an LSTM to capture temporal information, and allocates a larger share of samples to training while emphasizing recent observations in validation. AlphaNet v3 expands the feature extraction layers with different lookback windows and replaces the LSTM with a GRU to reduce model parameters.

The cited backtests cover Chinese A-share, CSI 800, and CSI 500 universes from January 2011 through July 2020, with rebalancing every ten trading days. The summary reports improved RankIC and information ratios for v2 versus v1 in the broad market and CSI 800, and modestly higher metrics for v3 versus v2 in CSI 500. These are historical results from the summarized report, not evidence of future performance. The comparison also notes a trade-off: AlphaNet automates factor discovery and combination but is less interpretable and supports a narrower set of feature operations than genetic programming paired with random forests.

Key ideas

• AlphaNet learns stock-selection factors from price and volume data in an end-to-end model.
• Version 2 adds ratio features, uses an LSTM for temporal patterns, and changes the training and validation split.
• Version 3 uses different feature lookback windows and a GRU with fewer parameters.
• The summarized historical tests report stronger results for v2 in broad A-share and CSI 800 universes, and a modest v3 improvement in CSI 500.
• AlphaNet reduces manual factor maintenance but is less interpretable than genetic programming paired with random forests.

Tags

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