EcoXplain: An Interpretable Causal-Augmented Framework for Macroeconomic Forecasting

Jiaxin Han, Shuo Miao, Changwang Li, Xuanzheng Wang, Chengqi Yi

Keywords: time-series forecasting, dynamic causality, macroeconomics

TL;DR: EcoXplain is an interpretable spatio-temporal neural network for macroeconomic forecasting, integrating white-box economic decomposition with dynamic causal graph learning.

Abstract: The urgency of policy response demands not only accurate predictions but also a deep understanding of causal mechanisms, which has become increasingly challenging as structural relationships between economic variables evolve over time. Existing machine learning approaches often function as black boxes, achieving high predictive accuracy but providing little interpretability, while traditional structural models struggle to adapt to fast-moving causal dynamics. To address these challenges, we propose EcoXplain, an interpretable dynamic causality augmented spatio-temporal graph neural network architecture specifically designed for low-frequency macroeconomic data. EcoXplain combines interpretable white-box decomposition with a Dynamic Spatio-Temporal Graph Neural Network (DSTGNN), which integrates short-term inferred dynamic causal relationships with prediction-driven adaptive adjacency matrices that capture evolving relationships between macroeconomic variables. Empirical evaluations on datasets from four major economies (China, Japan, the US, and the UK) show that EcoXplain significantly outperforms both white-box methods and state-of-the-art black-box machine learning baselines, reducing forecasting error (MAPE) by up to 88.98% and 68.07\%, respectively. Beyond predictive gains, EcoXplain uncovers meaningful causal pathways that provide policymakers and economists with a deeper understanding of how economic forces interact, offering valuable guidance for timely and effective decision-making.

Primary Area: learning on time series and dynamical systems

Submission Number: 7913