TL;DR
This paper introduces HATR, a novel deep learning model that tackles the challenge of stock trend prediction by simultaneously modeling temporal patterns within individual stocks and their interdependencies. HATR achieves state-of-the-art results by employing a hierarchical temporal module with multi-scale convolutions and a dual attention mechanism, coupled with a multi-graph relational module that learns adaptive stock relationships.
What Problem Does This Paper Address?
Predicting stock market trends is notoriously difficult due to the complex, non-stationary dynamics and intricate relationships between companies. Existing methods often treat stocks in isolation or rely on fixed, predefined relationships, failing to capture the nuanced, evolving interdependencies that influence market movements. Furthermore, traditional time-series models struggle to effectively capture both short-term volatility patterns and long-term trends within a stock's historical data. This paper aims to bridge this gap by proposing a model that can learn these temporal and relational dynamics in a hierarchical and adaptive manner.
Methods
The proposed Hierarchical Adaptive Temporal-Relational Network (HATR) consists of two main modules:
- Multi-Scale Temporal Representation Module: This module extracts volatility patterns from a stock's historical data. It uses:
- Stacked Dilated Causal Convolutions: These layers capture local context and gradually expand receptive fields across different time spans without a linear increase in parameters.
- Gated Linear Units (GLU): These are used to parcel convolutional operations, enabling both linear and non-linear information flow.
- Hawkes Process and Target-Specific Dual Attention: This mechanism identifies salient temporal points and scales, customizing stock representations based on individual stock traits and historical event impacts.
- Multi-Graph Stock Interrelation Module: This module captures dependencies among stocks. It constructs relationships from:
- Industry Graph (G1): Based on industry classification.
- Topicality Graph (GT): Derived from online discussions and co-mentions.
- Self-Adaptive Graph (Gapt): An automatically learned graph that discovers hidden dependencies without relying on prior domain knowledge.
- Diffusion Graph Convolution: This technique is used to propagate information across these multiple graph structures, allowing for a more comprehensive understanding of inter-stock relationships.
These modules are integrated into a unified, end-to-end framework for stock trend prediction.
Datasets
The paper evaluates HATR on three real-world stock market datasets:
- CSI (China): Comprises 300 stocks from the CSI-300 Composite Index.
- SPX (US): Includes 485 stocks from the S&P-500 Composite Index.
- Topix (Japan): Features 95 stocks from the TOPIX-100 Index.
The data spans from 2015 to 2020, with daily quote data (open, high, low, close prices, and trading volume) normalized. The training, validation, and test sets are split chronologically.
Key Insights and Ideas
- Hierarchical Temporal Modeling is Crucial: The paper demonstrates that capturing stock evolution at multiple temporal scales (short-term volatility and long-term trends) is essential for accurate prediction.
- Adaptive Relationships Outperform Fixed Ones: The self-adaptive graph learning mechanism significantly improves performance, suggesting that dynamically discovering stock interdependencies is more effective than relying solely on predefined industry or topicality links.
- Dual Attention Enhances Temporal Feature Extraction: The combination of Hawkes process for time-wise attention and target-specific query for scale-wise attention allows the model to focus on the most relevant historical information for each stock.
- Multi-Graph Fusion Captures Diverse Dependencies: Integrating information from industry, topicality, and adaptive graphs provides a richer representation of stock relationships than any single source alone.
Key Results
HATR achieves state-of-the-art performance across all three datasets, outperforming various baseline models, including traditional machine learning methods (SVM, RF), RNN-based models (DA-RNN, SFM, TPA-LSTM, TGC), and CNN-based models (InceptionTime).
- HATR demonstrates significant improvements in ACC, AUC, F1-score, and MCC metrics.
- The model is considerably faster than many deep learning baselines, being approximately 6.4x faster than DA-RNN and 18.2x faster than InceptionTime in training.
- Ablation studies confirm the effectiveness of both the temporal and relational modules, with HATR-MT (temporal only) and HATR-MR (relational only) showing performance degradations compared to the full HATR model.
Limitations and Future Work
- Limitations: While HATR performs well, the paper notes that the optimal setting for diffusion steps in the graph convolution is 2, suggesting that deeper graph interactions might lead to over-smoothing. The performance also declines with very high dimensions for ID embeddings, indicating a risk of overfitting.
- Future Work: The authors suggest exploring dynamic heterogeneous stock graphs that fuse multi-source information, such as news events, to further enhance the learning of time-evolving market dependencies.
Summary updated