From Causal Discovery to Dynamic Causal Inference in Neural Time Series

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

Dynamic Causal Network Autoregression (DCNAR) is a two-stage neural causal modeling framework for dynamic scientific systems with unknown or evolving causal networks. It addresses the limitation of existing approaches that assume a priori knowledge of causal structure. Stage one employs a neural autoregressive causal discovery model to learn a sparse directed network from multivariate time series. Stage two uses this network as a structural prior for time-varying neural network autoregression, enabling dynamic causal influence estimation. Evaluated via behavioral diagnostics like causal necessity and temporal stability, DCNAR demonstrates superior stability and interpretability. Experiments on multi-country panel time-series data (139 countries over 35 years, and 89 countries over 75 years) show DCNAR yields more stable, interpretable, and theoretically meaningful dynamic causal inferences than alternatives, even with comparable forecasting performance and near-nominal 90% prediction interval coverage.

Key takeaway

For research scientists analyzing complex adaptive systems, DCNAR provides a robust framework for dynamic causal inference when network structure is uncertain. You should adopt behavioral diagnostics like impulse responses and counterfactual simulations to validate your models' scientific utility, moving beyond prediction-centric evaluation. This approach helps you prioritize interpretable, stable causal behavior, avoiding the pathological dynamics often seen in purely predictive models, even if it means modest trade-offs in predictive optimality.

Key insights

DCNAR integrates causal discovery with dynamic inference for time-varying systems lacking known causal structure.

Principles

• Scientific models require behavioral diagnostics beyond predictive accuracy.
• Treat causal structure as a learned, testable prior, not a fixed assumption.
• Evaluate dynamic causal models by impulse responses and counterfactual stability.

Method

First, a neural additive vector autoregression (NAVAR) infers a sparse directed network; then, this network constrains a time-varying network autoregression (tvNAR) model for dynamic causal influence estimation.

In practice

• Apply DCNAR to analyze dynamic causal influence in complex adaptive systems.
• Assess model plausibility using impulse response functions and counterfactual trajectories.
• Utilize DCNAR for panel time-series data with unknown or evolving causal structures.

Topics

• Causal Discovery
• Dynamic Causal Inference
• Neural Time Series
• Network Autoregression
• Impulse Response Analysis
• Counterfactual Simulation

Best for: AI Scientist, Research Scientist

Related on AIssential

• Benchmarking Counterfactual Prediction in Epidemic Time Series with Time-Varying Interventions — The lack of realistic benchmarks with ground-truth counterfactuals constrains time-series causal inference progress.
• Causal Learning with the Invariance Principle — Two auxiliary environments suffice for causal graph and counterfactual inference in acyclic, invertible SCMs.
• Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time — Observability of latent states is crucial for identifying dynamic treatment effects in continuous-time causal forecasting with hidden confounders.
• Bridge: Basis-Driven Causal Inference Marries VFMs for Domain Generalization — The "Bridge" framework improves object detection domain generalization by using causal inference to block confounders.
• Test Time Training for Supervised Causal Learning — TTT-SCL improves Supervised Causal Learning's real-world applicability by dynamically generating test-aligned training sets.
• From Chaos to Causality: Debugging Multi-Agent Systems — Causal inference, using episode graphs and counterfactual replay, can debug multi-agent systems by identifying root causes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.