Introduction
Predicting stock market trends is a notoriously difficult task, often likened to a "random walk." However, recent research from Peking University and Mizuho Securities suggests that significant patterns exist in the "overnight gap"—the price movement between a market's close and its subsequent opening. Unlike intraday trading, where news is rapidly absorbed, the overnight period allows for a concentrated reaction to after-hours information. This paper introduces LSTM-RGCN, a graph-based neural network designed to predict these overnight movements by leveraging both financial news and the inherent relational structure of the stock market.
Methodology
The core innovation of this work is treating the stock market as a relational graph rather than a collection of isolated entities. The proposed LSTM Relational Graph Convolutional Network (LSTM-RGCN) architecture consists of four primary stages:
- Stock Correlation Graph: A graph is constructed where nodes represent individual stocks and edges denote historical price correlations (positive or negative).
- Node News Encoder: For stocks associated with news headlines, an LSTM with an attention mechanism encodes the text into a vector, using the stock’s embedding as a query to highlight relevant financial terms.
- Graph Encoder (RGCN): This component facilitates information propagation between connected stocks. To mitigate "over-smoothing," an LSTM gating mechanism is integrated between graph layers to control the flow of information.
- Global Node: A master node connected to all stocks tracks broader market trends, accounting for systemic factors that influence the entire index.
Experimental Setup
The researchers evaluated their model using a newly curated dataset comprising:
- Reuters Financial News: Headlines spanning 2013 to 2018.
- Tokyo Stock Exchange (TSE) Data: Price movements for the TPX500 and TPX100 indices.
- Overnight Metric: A "valid movement" was defined as a price change exceeding 0.5 times the hourly standard deviation, filtering out minor statistical noise.
Results and Discussion
The LSTM-RGCN model demonstrated superior performance over strong baselines, including Transformers and Hierarchical Attention Networks (HAN).
Impact of Graph Relations
A critical finding was the significant boost in accuracy provided by the graph structure. By modeling inter-stock dependencies, the "proposal" model outperformed a version of itself that treated stocks independently ("w/o graph").
Associative Inference and Zero-News Prediction
The graph-based approach enables "associative inference," where news affecting one company can help predict the movement of a correlated peer. In tests involving stocks with no direct news coverage, the model achieved 52–57% accuracy—significantly higher than a random baseline (50%)—demonstrating the power of information propagation across the market graph.
Challenges with Earnings Data
The study noted that "earnings" news remains a difficult category to model. Unlike general "affairs" news, which often has a clear sentiment, earnings reports require complex numerical reasoning (e.g., comparing actual results against market expectations), which standard text encoders are currently less equipped to handle.
Conclusion
The LSTM-RGCN model represents a significant step forward in utilizing relational data for financial forecasting. By focusing on the overnight gap and leveraging inter-stock correlations, the model captures market dynamics that traditional models overlook. Future research may focus on integrating more diverse relationship types, such as supply chain links, and enhancing the text encoder's capacity for numerical reasoning.
Summary updated