aiqbench analysis
This paper advances causal inference in time-series analysis by integrating Granger causality with Bayesian networks to better distinguish direct effects from indirect dependencies. By validating these methods on complex dynamical systems and biological data, the work offers a rigorous framework for interpreting temporal prediction as genuine causal evidence rather than mere correlation.
Re-examining Granger Causality with Causal Bayesian Networks and Reichenbachs Principles
arXiv:2501.02672v4 Announce Type: replace-cross Abstract: Granger causality (GC) is widely used to infer directed relationships in time-series data. However, its predictive criterion does not by itself distinguish direct causal effects from dependencies induced by common causes, indirect paths, collider conditioning, or model misspecification. We revisit this limitation by interpreting bivariate and multivariate GC through causal Bayesian networks and Reichenbachs common cause principles. Under explicit graphical assumptions, bivariate GC provides a marginal dependence check, while multivariate GC tests whether the same association persists after conditioning on relevant histories. This view motivates causalised Granger causality (c-GC), which combines the two decisions, and c-GC*, a more conservative variant with a richer conditioning set. We validate both methods on synthetic dynamical systems, established time-series causal discovery benchmarks, Sachs protein-signalling data, and Lorenz-96 simulations. The results show that the proposed criteria recover plausible causal structure in settings with delayed effects, cycles, bidirectional links, and nonlinear or noisy dynamics. The framework clarifies how GC-style inference can be given a causal interpretation without treating temporal prediction alone as sufficient evidence of causation.
Abstract:Granger causality (GC) is widely used to infer directed relationships in time-series data. However, its predictive criterion does not by itself distinguish direct causal effects from dependencies induced by common causes, indirect paths, collider conditioning, or model misspecification. We revisit this limitation by interpreting bivariate and multivariate GC through causal Bayesian networks and Reichenbachs common cause principles. Under explicit graphical assumptions, bivariate GC provides a marginal dependence check, while multivariate GC tests whether the same association persists after conditioning on relevant histories. This view motivates causalised Granger causality (c-GC), which combines the two decisions, and c-GC*, a more conservative variant with a richer conditioning set. We validate both methods on synthetic dynamical systems, established time-series causal discovery benchmarks, Sachs protein-signalling data, and Lorenz-96 simulations. The results show that the proposed criteria recover plausible causal structure in settings with delayed effects, cycles, bidirectional links, and nonlinear or noisy dynamics. The framework clarifies how GC-style inference can be given a causal interpretation without treating temporal prediction alone as sufficient evidence of causation.
Machine Learning (stat.ML); Machine Learning (cs.LG); Econometrics (econ.EM); Methodology (stat.ME)
arXiv:2501.02672 [stat.ML]
(or arXiv:2501.02672v4 [stat.ML] for this version)
https://doi.org/10.1142/S0219525926500037 DOI(s) linking to related resources
Full text under the publisher's open licence · 194 words
Related coverage
Bayesian Networks with Latent Time Embedding for Stage-Aware Causal Modeling of Alzheimer's Disease Progression
arXiv · Jun 16, 2026
A Gradient-based Causal Discovery Framework with Applications to Complex Industrial Processes
arXiv · Jun 17, 2026
Amortized Bayesian Causal Discovery of Extended Factor Graphs
arXiv · Jul 28, 2026
PIRAMID: Progress and Plans
LessWrong · Jul 27, 2026