Using Graph Attention Networks to Predict Short-Horizon Stock Returns
Summary
This article introduces graph neural networks and focuses on graph attention networks (GATs) for financial time-series prediction. It explains that GATs use attention to weight neighboring nodes' features, allowing the weights to depend on node data rather than being fixed by the graph structure. The article describes this as a potential advantage for generalization and parallel computation, while noting that the approach may miss known relationships between securities unless those relationships are incorporated explicitly.
The proposed stock-selection experiment predicts five-day returns from seven price and volume inputs transformed into 98 factors. It describes training on two years of daily market data and testing on the following year, with a separate rolling evaluation design. Preprocessing fills missing values with zero, applies cross-sectional Z-score scaling to factors and labels, clips values, and uses a five-step input window. The article gives no completed backtest results or detailed model settings; its training section is unfinished. It also inconsistently refers to TabNet in the experiment description, so the reported setup should be treated cautiously.
Key ideas
- GATs use attention weights to aggregate neighboring node features.
- The article argues that feature-based attention may generalize better across graph structures than graph-dependent aggregation.
- The proposed stock model predicts five-day returns using 98 factors derived from seven daily market inputs.
- The described preprocessing includes missing-value filling, cross-sectional scaling, clipping, and a five-step time window.
- The document provides no completed backtest results, and its experiment section contains a model-name inconsistency.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.