How Can Supply Chains Forecast Stock Futures?

Imagine reading a stock's future not only by looking at its chart, but also by seeing who supplies its parts and who buys them. That's exactly what the new Full‑State Graph Convolutional LSTM (FS‑GCLSTM) model does.

Researchers turned business relationships—suppliers, customers, partners—into a network where each node is a company and each edge a commercial link. Then they fed this network into an LSTM, a type of AI that excels at understanding time sequences like stock prices.

The breakthrough is that the model doesn't use the network just once; it "convolves" it at every LSTM step: input, hidden state, and cell state. Thus, supply‑chain information shapes every update, making predictions more aware of real‑world context.

Tested on two major indices—Eurostoxx 600 and S&P 500—the FS‑GCLSTM isn’t always the most accurate in average error, but it outperforms other models in portfolio performance: higher annual returns and better risk‑adjusted metrics (Sharpe and Sortino). Gains are clearer in markets with denser networks, like Eurostoxx 600, and the model stays stable even when the input sequence length changes.

In short, merging supply‑chain data with temporal graph neural networks opens a fresh way to see how companies affect each other and, consequently, how their stocks move.

The FS‑GCLSTM merges two advanced architectures: Graph Convolutional Networks (GCN) and Long Short‑Term Memory (LSTM). GCNs operate on graphs, propagating information from a node to its neighbors via a normalized adjacency matrix \(\hat{A}=D^{-1/2}(A+I)D^{-1/2}\), where \(D\) is the degree matrix and \(I\) the identity, ensuring numerical stability. In FS‑GCLSTM, this operation is applied not only to the input \(X_t\) but also to the hidden state \(h_{t-1}\) and cell state \(c_{t-1}\). The result is a set of updates: \[\tilde{X}_t = \sigma(\text{GCN}(X_t)),\] \[\tilde{h}_{t-1} = \sigma(\text{GCN}(h_{t-1})),\] \[\tilde{c}_{t-1} = \sigma(\text{GCN}(c_{t-1})).\] These convoluted terms feed the LSTM gates (input, forget, output) with spatial information, allowing the model to capture temporal and topological dependencies simultaneously.

From a data perspective, the value‑chain network is built using B2B relationships supplied by LSEG, including transaction volumes, trade frequency, and industry sector. Node features comprise historical prices, volatility, and fundamental indicators (P/E, ROE). Feature normalization is crucial to prevent variables with different scales from dominating signal propagation.

Training uses a Mean Squared Error (MSE) loss between predicted and actual returns, but final evaluation relies on portfolio performance metrics: annualized return, Sharpe ratio (μ/σ), and Sortino ratio (μ/σ_down). Results show that, despite a slightly higher MSE compared to pure LSTM models, FS‑GCLSTM generates more robust portfolios thanks to its ability to anticipate supply‑chain shocks invisible to time‑series‑only models.

A sensitivity analysis revealed that graph density (average number of connections per node) directly impacts predictive power: denser graphs provide more propagation signals, enhancing the model's generalization. Moreover, robustness tests with varying input sequence lengths (10 to 60 days) confirmed model stability, with performance variations under 3%. This suggests that integrating supply‑chain data with temporal graph neural networks is a promising direction for computational finance.