Post-hoc Calibration of Artificial Intelligence Crop-Yield Ensembles Under Regime Shifts: Identification Bounds and Smallholder Insurance Basis-Risk Equity

Post-hoc calibration of artificial intelligence crop-yield ensembles is widely proposed as a remedy for the distributional instability that undermines index insurance in smallholder settings. The problem is formalised as a partial-identification question: the calibration map from ensemble predictions to yield realisations is not point-identified when the joint distribution of predictors and outcomes shifts between training and deployment, but it may be bounded. A calibration-identification bound is derived from the worst-case divergence between source and target predictor distributions, and the bound is shown to widen monotonically with the magnitude of the shift. The analysis then connects the bound to basis-risk equity, defined as the dispersion of the gap between insurance payouts and realised losses across smallholder subgroups. A key result is that post-hoc calibration can reduce aggregate basis risk while increasing subgroup dispersion, because the calibration map is estimated on a pooled sample whose subgroup composition differs from the deployment population. The argument is methodological and theoretical; no empirical calibration is performed, and the results are propositions about identification rather than estimates from data.