Abstract
This study proposes a novel framework that integrates causal regularization into hybrid Graph Attention Network and Long Short-Term Memory models for supply chain demand forecasting. We incorporate causal structures discovered via DYNOTEARS (Pamfil et al., 2020) to guide the attention mechanisms of spatiotemporal neural networks, enabling the model to learn forecasting patterns aligned with underlying causal dependencies. Through comprehensive experimentation on the SupplyGraph benchmark dataset (Wasi et al., 2024), we demonstrate that explicitly modeling causal relationships substantially improves forecasting accuracy. The baseline hybrid GAT-LSTM model achieves an RMSE of 1.124 with an \(R^2\) of 0.364, while our causally regularized variant reduces RMSE to 0.986 and improves \(R^2\) to 0.511 (from \(R^2 = 0.364\) to \(R^2 = 0.511\) , a 40.19% relative improvement), representing a 12.27% improvement in prediction accuracy. Systematic ablation studies confirm that the performance gains arise specifically from the incorporation of DYNOTEARS-discovered causal structure rather than arbitrary graph regularization. These results demonstrate that principled integration of causal discovery methods can enhance both the accuracy and interpretability of spatiotemporal neural forecasting models, offering a promising direction for developing more robust and explainable supply chain analytics systems.
Graphical abstract
