A new paper by Christopher Lohse and CERTAIN researcher Jonas Wahl in Transactions on Machine Learning Research extends varsortability (a measure of the agreement between the order of increasing marginal variance and the causal order) and R²-sortability (sorting the variables by increasing R2 yields an ordering close to a causal order) to time series—and shows why that matters for evaluating causal discovery.
The authors adapt sortability metrics to time-series and test them on simulated and real datasets (climate challenge data, river flows, Causal Chamber). The results show that some real-world sets show high varsortability, others do not. Whether scales carry causal signals depends on the domain of application.
Additionally, the performance of continuous, score-based methods (e.g., DYNOTEARS) correlates with varsortability, suggesting that benchmarks should report these assumptions, and assumptions like equal noise variance need scrutiny.