Graph Neural Networks for Stocks

Article BigQuant

Summary

The document introduces graph representations for modeling relationships among financial assets, arguing that stocks cannot be understood solely as independent observations. Graphs can encode different kinds of links, including sector membership, return correlation, and supply-chain connections, whose strengths and structures do not fit naturally into ordinary spatial convolution. Graph neural network layers can pass information across these links so a model can use each stock’s broader market context when forecasting returns.

The paper overview reports experiments with three graph neural network layers for stock return prediction and portfolio construction. It says graph layers stabilized forecasts relative to conventional methods such as LSTMs, reduced trading costs, and filtered high-frequency signals. It also reports that supply-chain links were more informative in 2021 than sector or correlation graphs. These claims are summarized without experimental details, sample definitions, or performance metrics, so the excerpt does not establish how robust the findings are across periods or markets.

Key ideas

Graphs can represent multiple types of relationships among stocks, including sector, correlation, and supply-chain links.
Graph neural layers propagate information between connected assets to contextualize return forecasts.
The paper evaluates three graph neural network layers and uses their predictions to construct portfolios.
The overview reports forecast stabilization, lower trading costs, and filtering of high-frequency signals.
It says supply-chain information was more useful than sector or correlation links in 2021, but gives no detailed validation in the excerpt.

Using Graph Neural Networks to Model Relationships Between Stocks

Summary

Key ideas

Tags