[Paper Review] Exploiting Supply Chain Interdependencies for Stock Return Prediction: A Full-State Graph Convolutional LSTM

Chang Liu|arXiv (Cornell University)|Mar 7, 2023

Stock Market Forecasting MethodsDecision Sciences3 citations

TL;DR

This paper proposes an LSTM-GCN model that leverages supply chain interdependencies through a graph-structured value chain to improve stock return prediction. By combining graph convolutional networks to capture spatial dependencies among companies and LSTMs to model temporal dynamics, the method achieves superior prediction accuracy and generates significant excess returns on both Eurostoxx 600 and S&P 500 datasets.

ABSTRACT

Stock return prediction is fundamental to financial decision-making, yet traditional time series models fail to capture the complex interdependencies between companies in modern markets. We propose the Full-State Graph Convolutional LSTM (FS-GCLSTM), a novel temporal graph neural network that incorporates value-chain relationships to enhance stock return forecasting. Our approach features two key innovations: First, we represent inter-firm dependencies through value-chain networks, where nodes correspond to companies and edges capture supplier-customer relationships, enabling the model to leverage information beyond historical price data. Second, FS-GCLSTM applies graph convolutions to all LSTM components - current input features, previous hidden states, and cell states - ensuring that spatial information from the value-chain network influences every aspect of the temporal update mechanism. We evaluate FS-GCLSTM on Eurostoxx 600 and S&P 500 datasets using LSEG value-chain data. While not achieving the lowest traditional prediction errors, FS-GCLSTM consistently delivers superior portfolio performance, attaining the highest annualized returns, Sharpe ratios, and Sortino ratios across both markets. Performance gains are more pronounced in the denser Eurostoxx 600 network, and robustness tests confirm stability across different input sequence lengths, demonstrating the practical value of integrating value-chain data with temporal graph neural networks.

Motivation & Objective

• To improve stock return prediction by incorporating supply chain interdependencies beyond traditional price data.
• To model spatial dependencies among companies using graph convolutional networks (GCN) based on supplier-customer relationships.
• To capture temporal dynamics in stock returns using long short-term memory (LSTM) networks.
• To evaluate the model’s performance against baseline models in terms of prediction error and portfolio profitability.
• To assess the robustness of the model’s predictive power across varying rolling window lengths.

Proposed method

• Construct a graph where nodes represent companies and edges represent supplier-customer relationships from value chain data.
• Use node features derived from historical stock returns to represent each company in the graph.
• Apply graph convolutional networks (GCN) to extract spatial dependencies and topological structure from the graph.
• Integrate LSTM networks to model temporal evolution of node features and capture long-term dependencies in return sequences.
• Combine GCN and LSTM in a unified architecture (LSTM-GCN) to jointly model spatial and temporal patterns.
• Train the end-to-end model using mean absolute error (MAE) and mean squared error (MSE) as loss functions.

Experimental results

Research questions

• RQ1 Can modeling supply chain interdependencies through a graph structure improve stock return prediction accuracy?
• RQ2 Does the integration of GCN and LSTM outperform standalone models in capturing both spatial and temporal dependencies in financial data?
• RQ3 Can the LSTM-GCN model generate economically significant excess returns in a simulated market-neutral trading strategy?
• RQ4 Is the model’s predictive performance robust across different market indices and rolling window lengths?
• RQ5 Does the model’s performance significantly exceed that of using a simple mean forecast, as validated by statistical tests?

Key findings

• The LSTM-GCN model achieved the lowest MAE (7.1993 × 10⁻⁴) and MSE (1.6607 × 10⁻²) on the Eurostoxx 600 dataset, outperforming all baselines.
• On the S&P 500, the LSTM-GCN model recorded the highest annualized return (5.8904%) and Sharpe ratio (0.5345), indicating superior risk-adjusted performance.
• The model generated the highest cumulative return and Sortino ratio in backtested market-neutral strategies, demonstrating economic significance.
• Robustness testing showed that for Eurostoxx 600, R² was significantly greater than zero for rolling windows of 50, 60, 70, and 80 days (p < 0.001).
• For S&P 500, R² was significantly greater than zero for 50 and 60-day rolling windows, with p-values of 2.569 × 10⁻⁷ and 7.286 × 10⁻¹⁰, respectively.
• The model’s performance was significantly better than a mean-forecast baseline, confirming its predictive power beyond random chance.