Strategies for spatial leave-one-out cross-validation

The purpose of the talk is to discuss spatially adapted cross-validation methods that maintain sufficient separation between training and validation sets, thus providing more accurate estimates of model risk. We begin by reviewing various spatial cross-validation techniques, including spatial blocked cross-validation and spatial leave-one-out, under scenarios of low to strong spatial dependence. We then propose a practical framework for determining an optimal “buffer size” for spatial leave-one-out that reduces autocorrelation between training and validation subsets. This framework is further enhanced by a parametric bootstrap approach designed to approximate the true risk in single-realization settings. Simulation experiments confirm that these methods effectively capture the underlying spatial structure, leading to more reliable risk estimation.