HiBaNG tackles low-data Granger
A new paper on Hierarchical Bayesian Nonparametric Granger causal discovery (HiBaNG) promises better causal inference in low-data regimes — directly relevant for macro-finance and short-sample cross-sections. The approach could be applied when standard Granger tests lack power, e.g., small-country macro panels or event-limited windows. (x.com)
HiBaNG was posted to OpenReview on Oct. 9, 2025 and the forum entry was last modified on Jan. 18, 2026 under submission number 6167, with Emmanuel Bengio listed as the assigned action editor. (openreview.net) The author list on the official bibliographic record names He Zhao, Vassili Kitsios, Terence O’Kane and Edwin V. Bonilla. (jmlr.org) The paper defines a hierarchical, factorized prior over binary Granger-causal graphs that encodes structured sparsity and reports interpretable, uncertainty-aware posterior statements about graph edges. (openreview.net) Inference is implemented with a Gibbs sampler that the authors say exploits conjugacy and data augmentation to make posterior estimation tractable and scalable. (openreview.net) Empirical results reported in the manuscript compare HiBaNG against classical VAR and recent deep-VAR baselines on synthetic, semi-synthetic and real-world climate datasets, with the authors claiming improvements in accuracy and in uncertainty calibration. (openreview.net) This work follows prior related publications from the same team, including a 2024 arXiv paper on Bayesian VARs with factorised Granger-causal graphs that shares methodological themes and authorship. (arxiv.org)
