A tiered spatial validation workflow for geographic prediction models.

Geographically indexed observations violate the exchangeability assumed by ordinary random validation. We present a tiered spatial validation workflow that separates interpolation within an observed geographic mixture from transfer to a distinct area. The workflow combines random folds, a size matched random control, coordinate clusters, spectral communities of a nearest neighbor graph, administrative holdouts, and an external regional stress test. It reports pooled R squared and root mean squared error, repeated seed variability, spatial autocorrelation, and sensitivity to block scale. The workflow is dataset agnostic and is demonstrated on 211 geolocated Baltimore house sales predicted from structural attributes. Random validation exceeded coordinate cluster and graph partition validation, and the size matched control ruled out a training size artifact. Administrative and external region holdouts reduced apparent performance further, and coordinate cluster validation was closer to true external region transfer than random validation but remained optimistic. The workflow is intended for geographic machine learning studies that need to match performance claims to a stated deployment setting. •The workflow distinguishes interpolation, separation within an observed network, and external geographic transfer. •Matched size controls, repeated partitions, autocorrelation diagnostics, and regional holdouts make the validation gap interpretable. •A dataset agnostic reference implementation is demonstrated on a public housing dataset unrelated to the motivating application.