Causal Discovery Framework for Non-Stationary Time Series

A novel approach for uncovering causal relationships in multivariate time series has been developed to tackle issues related to non-stationarity and autocorrelation. Named Decomposition-based Causal Discovery (DCD), this technique dissects each time series into its trend, seasonal, and residual parts. Stationarity tests are utilized for trend components, kernel-based dependence measures for seasonal components, and constraint-based causal discovery for residual components. Subsequently, the graphs at the component level are merged into a cohesive multi-scale causal framework. This method effectively distinguishes between long- and short-range causal influences while minimizing false connections. The findings are available on arXiv with the identifier 2602.01433.