Towards E-Values for Causal Discovery at Earth-Observation Scale

Invariant causal prediction (ICP) offers a principled route to causal discovery from observational data, and is well suited to Earth Observation (EO) tasks, where the same processes are observed under many different conditions; yet its guarantee comes from testing every possible combination of candidate drivers, which at global scale is unaffordable. Rather than accept the loss of guarantee that comes with heuristic search, we replace the statistical test inside ICP with a universal-inference e-value, which requires no resampling. We examine this construction on the SeasFire datacube, testing eleven ocean-climate indices as candidate drivers of global wildfire, and compare it against the greedy ICP baseline it replaces. We find that the e-value construction makes the method both more stable across subsamples and dramatically cheaper to compute. Our results motivate extending this approach to the full subset search, and to the wider set of candidate drivers that EO data offers.