Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Chang Liu

TL;DR

The paper tackles stock return prediction by incorporating inter-firm value-chain structure through a Full-State Graph Convolutional LSTM (FS-GCLSTM). It advances the method by applying graph convolutions to all LSTM inputs (current features, previous hidden, and cell states), enabling spatial information from supplier–customer networks to influence temporal updates. Evaluated on Eurostoxx 600 and S&P 500 with LSEG value-chain data, FS-GCLSTM achieves superior portfolio-level performance (higher annualized returns, Sharpe, and Sortino) despite not always having the lowest traditional prediction errors, with gains amplified in denser networks. The work demonstrates the practical value of merging value-chain data with temporal graph networks for investment decision-making and outlines avenues for richer, multi-modal graph modeling.

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.

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

TL;DR

The paper tackles stock return prediction by incorporating inter-firm value-chain structure through a Full-State Graph Convolutional LSTM (FS-GCLSTM). It advances the method by applying graph convolutions to all LSTM inputs (current features, previous hidden, and cell states), enabling spatial information from supplier–customer networks to influence temporal updates. Evaluated on Eurostoxx 600 and S&P 500 with LSEG value-chain data, FS-GCLSTM achieves superior portfolio-level performance (higher annualized returns, Sharpe, and Sortino) despite not always having the lowest traditional prediction errors, with gains amplified in denser networks. The work demonstrates the practical value of merging value-chain data with temporal graph networks for investment decision-making and outlines avenues for richer, multi-modal graph modeling.

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.
Paper Structure (13 sections, 3 equations, 7 figures, 5 tables)

This paper contains 13 sections, 3 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Previous Research
  3. The FS-GCLSTM Model
  4. Graph Convolutional Layer.
  5. FS-GCLSTM Cell.
  6. Model Architecture.
  7. Empirical Analysis
  8. Data
  9. Preprocessing
  10. Model Set Up
  11. Baseline Models
  12. Evaluation Metrics
  13. Conclusion

Figures (7)

  • Figure 1: Schematic representation of the FS-GCLSTM cell. The diagram illustrates how GCN layers process the previous hidden state $\mathbf{h}_{t-1}$, previous cell state $\mathbf{c}_{t-1}$, and current input features $\mathbf{X}_t$ before they enter the LSTM gates for the update computations.
  • Figure 2: Architecture of the FS-GCLSTM model. Left: Rolling windows of historical stock price data (d trading days) form node features for value chain graphs at each time step. Center: Three stacked FS-GCLSTM cells process the temporal sequence, with each cell applying GCN transformations to all LSTM inputs. Right: Final hidden states are concatenated, flattened, and passed through an MLP to predict next-day returns for selected stocks. The number of output predictions $N_{pred}$ may be different than the total number of nodes in the graph.
  • Figure 3: Value chain graphs of Eurostoxx600 and S&P 500 datasets. Nodes are the listed companies. Red nodes denote companies that are constituents of the corresponding indexes. Edges represent the supplier/customer relationships between two entities. In this layout, components less connected to other components are placed in peripheral space. The graph of Eurostoxx 600 looks denser, although the total number of edges in both graphs is comparable. This is due to the fact that the connections in S&P value chain data are more concentrated on fewer nodes in the inner space.
  • Figure 4: Cumulative portfolio performance for Eurostoxx 600 using daily rebalanced, equal-weighted, long-only strategies. Portfolio allocation is restricted to current index constituents from the 19 European countries. FS-GCLSTM consistently outperforms all baseline models (ARIMA, FCL, LSTM, GConvGRU) and the constant-weights benchmark throughout the evaluation period.
  • Figure 5: Cumulative portfolio performance for S&P 500 using daily rebalanced, equal-weighted, long-only strategies. Portfolio allocation is restricted to current U.S.-listed index constituents. FS-GCLSTM achieves the highest final portfolio value and superior risk-adjusted returns, exceeding all baseline models and the constant-weights benchmark.
  • ...and 2 more figures