Causal Discovery from Time Series Data Using Neural Differential Equations

Identifying causal relationships from observational time series data is a fundamental challenge in scientific discovery, complicated by latent confounders, measurement noise, and non-linear dynamics. We introduce CausalNDE, a framework that combines neural differential equations with score-based causal discovery to infer causal graphs from multivariate time series. CausalNDE models continuous-time system dynamics as neural ODEs and employs an interventional score criterion to distinguish causal from spurious correlations. Applied to synthetic benchmarks and real-world datasets from neuroscience (EEG), climate science (climate indices), and economics (financial markets), CausalNDE identifies true causal graphs with F1 scores 15–23% higher than Granger causality baselines, while providing interpretable causal strength estimates.