AlphaNet Neural Stock Factor Discovery

Article BigQuant

Summary

This edited seminar transcript surveys a Chinese quantitative research program on deep learning, then focuses on AlphaNet, a neural approach to stock factor discovery. It frames a multi-factor process as factor generation, factor combination, and portfolio optimization. Earlier approaches replace linear combination with tree models or use genetic programming to generate factors. AlphaNet instead learns factor construction and combination together: custom operations capture relationships between arbitrary inputs and rolling statistics, then an LSTM predicts returns. The design aims to avoid convolutional filters whose results can depend on an arbitrary ordering of financial factors.

The transcript reports a 500-stock strategy with excess returns of about 20% versus a benchmark and an information ratio above 3, over a period beginning in 2011. It gives no detailed methodology, validation setup, transaction-cost treatment, or risk analysis, so those figures cannot establish robustness. The broader discussion also raises graph neural networks and GAN-generated market sequences, but the excerpt leaves these mostly as questions. It emphasizes overfitting and limited data as central challenges; portfolio optimization is not yet integrated into AlphaNet.

Key ideas

AlphaNet combines factor generation and factor synthesis in one neural network trained to predict stock returns.
Custom rolling operations are used to extract relationships between factors without relying on arbitrary factor ordering.
The described architecture sends engineered features into an LSTM, while portfolio optimization remains a separate step.
The transcript reports benchmark-relative results but omits enough validation and cost details to assess robustness.
The broader research agenda includes overfitting controls, synthetic data, graph networks, and market interpretation.

AlphaNet for Neural Stock Factor Discovery and Deep Learning in Quant Research

Summary

Key ideas

Tags