Introduction
Predicting stock market trends remains one of the most challenging tasks in financial forecasting due to high volatility and the intricate, non-linear dependencies between individual securities. While recent advancements have leveraged Graph Neural Networks (GNNs) to capture inter-stock relationships, many existing models suffer from three primary limitations: they ignore multi-scale relationships (focusing only on short-term dynamics), utilize simplistic integration methods (such as basic feature concatenation), and fail to model higher-order interactions within industry hierarchies.
To address these gaps, this paper introduces LSR-IGRU (Long Short-term Relationship and Improved GRU), a novel framework designed to capture both "slow" structural industry ties and "fast" dynamic price correlations, integrating them directly into the temporal learning process.
Methodology
The LSR-IGRU model is built upon three core components that allow it to process both temporal sequences and relational graphs simultaneously.
1. Long-term Relationship Modeling
The model constructs a hierarchical graph based on secondary industry classifications. By connecting individual stocks to their secondary industries, and those secondary industries to broader primary industries, the model captures "invariant" structural relationships. This hierarchy allows the model to learn how broad sector movements trickle down to specific stocks, providing a stable foundation for prediction that does not change on a daily basis.
2. Short-term Relationship Modeling
To capture rapid market shifts, the model utilizes overnight price fluctuations. By calculating the cosine similarity of overnight returns between stocks over a specific look-back period, LSR-IGRU builds a dynamic graph of stocks that are currently moving in tandem. This approach treats overnight price changes as a proxy for investor sentiment and information synergy, identifying short-term bonds that intra-day models often overlook.
3. The Improved GRU (IGRU)
The architectural centerpiece of the framework is the Improved GRU cell. In a standard GRU, the input at each time step is limited to the stock's own historical data. In contrast, the IGRU modifies the update gate, reset gate, and candidate hidden state to incorporate a concatenation of the stock’s temporal features, its long-term industry embedding, and its short-term price-action embedding. By using Graph Attention Networks (GATs) to derive these embeddings, the relational context modulates the temporal state at every step of the sequence, ensuring that "who a stock is related to" is inseparable from "what the stock did."
Experimental Evaluation
The model was rigorously evaluated using data from 2018 to 2023 across four major global indices:
- Chinese Market: CSI 300 and CSI 500.
- US Market: S&P 500 and NASDAQ 100.
The feature set included standard OHLCV data (Open, High, Low, Close, Volume), turnover rates, and industry codes. LSR-IGRU was benchmarked against 18 diverse baselines, including traditional RNNs (LSTM/GRU), Transformers, Reinforcement Learning models (PPO/SAC), and state-of-the-art graph-based models like THGNN and AlphaStock.
Key Results and Insights
LSR-IGRU demonstrated superior performance across all datasets, particularly in terms of Annualized Return Rate (ARR) and Information Ratio (IR):
- Superior Returns: On the NASDAQ 100, LSR-IGRU achieved an ARR of 83.5%, significantly outperforming the best graph-based baseline, DeepTrader (71.6%). On the S&P 500, it reached an ARR of 50.6%, nearly doubling the performance of standard deep learning architectures.
- Risk-Adjusted Performance: The model maintained an impressively low Excess Return Drawdown (approximately 5%) in backtests on the CSI 300 and 500, suggesting that the integration of long-term industry data provides a stabilizing effect during market volatility.
- Real-world Validation: Beyond academic benchmarks, the model was successfully integrated into the algorithmic trading system of Emoney Inc., a Chinese financial services provider, where it consistently outperformed existing production models.
Conclusion
The success of LSR-IGRU highlights the importance of multi-scale relational modeling in financial contexts. By treating industry structures and overnight price synergies as fundamental components of the temporal hidden state—rather than mere auxiliary features—the model achieves a more robust understanding of market dynamics.
Future research directions include accounting for transaction costs (slippage and taxes) to better reflect real-world net returns, exploring dynamic industry labels for companies that operate across multiple sectors, and integrating alternative data sources such as news sentiment or limit order book (LOB) data to further refine the model's predictive accuracy.
Summary updated