HiBaNG: Hierarchical Bayesian Nonparametric Granger Causal Discovery in Low-Data Regimes

Abstract

We present a principled probabilistic framework for discovering Granger causal relationships from multivariate time-series data in low-data regimes, where short sequences limit the applicability of modern deep learning approaches. While deep neural vector autoregressive (VAR) models perform well in high-data settings, they often struggle to generalize with limited samples and provide little insight into model uncertainty. To address these challenges, we introduce HiBaNG, a hierarchical Bayesian nonparametric framework for Granger causal discovery. HiBaNG places a hierarchical factorized prior over binary Granger causal graphs that encodes structured sparsity and enables interpretable, uncertainty-aware inference. We develop a tractable Gibbs sampling algorithm that exploits conjugacy and augmentation for scalable posterior estimation. Extensive experiments on synthetic, semi-synthetic, and real-world climate datasets demonstrate that HiBaNG consistently outperforms both classical and deep VAR baselines, achieving improved accuracy and calibrated uncertainty.