[Paper Review] HIST: A Graph-based Framework for Stock Trend Forecasting via Mining Concept-Oriented Shared Information

Wentao Xu, Liu Wei-qing+5 more|arXiv (Cornell University)|Oct 26, 2021

Stock Market Forecasting Methods43 references21 citations

TL;DR

HIST proposes a graph-based framework that dynamically models shared information from both predefined and hidden stock concepts to improve stock trend forecasting. By jointly learning dynamic concept representations and mining latent hidden concepts via a bipartite graph and residual architecture, HIST achieves a 120% cumulative return (70% better than CSI 300) in investment simulations from 2017–2020.

ABSTRACT

Stock trend forecasting, which forecasts stock prices' future trends, plays an essential role in investment. The stocks in a market can share information so that their stock prices are highly correlated. Several methods were recently proposed to mine the shared information through stock concepts (e.g., technology, Internet Retail) extracted from the Web to improve the forecasting results. However, previous work assumes the connections between stocks and concepts are stationary, and neglects the dynamic relevance between stocks and concepts, limiting the forecasting results. Moreover, existing methods overlook the invaluable shared information carried by hidden concepts, which measure stocks' commonness beyond the manually defined stock concepts. To overcome the shortcomings of previous work, we proposed a novel stock trend forecasting framework that can adequately mine the concept-oriented shared information from predefined concepts and hidden concepts. The proposed framework simultaneously utilize the stock's shared information and individual information to improve the stock trend forecasting performance. Experimental results on the real-world tasks demonstrate the efficiency of our framework on stock trend forecasting. The investment simulation shows that our framework can achieve a higher investment return than the baselines.

Motivation & Objective

• To address the limitation of static conceptual connections in existing graph-based stock forecasting models.
• To discover hidden concepts not captured by manually defined stock concepts, which may carry critical shared market information.
• To jointly model predefined and hidden concept representations for improved information propagation in stock trend forecasting.
• To integrate individual stock features with shared concept information for enhanced forecasting performance.
• To validate the framework’s effectiveness through real-world stock market experiments and investment simulation.

Proposed method

• Constructs a stock-concept bipartite graph to model relationships between stocks and both predefined and hidden concepts.
• Introduces a doubly residual architecture to extract and refine representations of hidden concepts from the bipartite graph.
• Learns dynamic relevance weights between stocks and concepts to enable context-aware information propagation.
• Uses a GRU-based module to extract temporal features from stock price and volume sequences.
• Combines individual stock features, predefined concept features, and hidden concept features via a gated fusion mechanism.
• Applies graph neural network layers to aggregate and propagate concept-oriented shared information across stocks.

Experimental results

Research questions

• RQ1 Can dynamically learned concept representations improve stock trend forecasting beyond static conceptual connections?
• RQ2 Can mining hidden concepts from stock data enhance forecasting performance compared to relying only on predefined concepts?
• RQ3 How do individual stock features compare in importance to shared concept information in forecasting?
• RQ4 Does the integration of predefined and hidden concepts lead to better investment performance in real-world simulations?
• RQ5 What kind of hidden concepts emerge from the model, and do they align with meaningful market events or sectors?

Key findings

• HIST achieved a 120% cumulative return from 2017 to 2020 on the CSI 300 index, outperforming the index by 70%.
• The investment simulation showed HIST outperformed state-of-the-art models including Transformer, ALSTM+TRA, and GATs.
• Ablation studies confirmed that removing either the predefined or hidden concept module degraded performance, proving both are essential.
• Removing the individual information module also reduced performance, indicating its critical role in forecasting.
• Visualizations of the hidden concept module revealed meaningful clusters, such as 'high-speed train' and 'jet fuel', aligning with real-world industrial groupings.
• The framework’s dynamic concept weighting mechanism effectively captured shifting market relevance, such as e-commerce’s rise during the pandemic.