Conformal prediction intervals for crop-yield forecasts under distribution shift: evaluating artificial intelligence calibration coverage and replanting-decision regret across smallholder advisories

Conformal prediction offers distribution-free coverage guarantees for crop-yield forecasts, but its validity depends on exchangeability between calibration and deployment data, an assumption routinely violated when smallholder advisories transfer models across seasons, agro-ecological zones or management regimes. The framework is then embedded in a decision-theoretic account of replanting advice, in which the cost of a missed replanting signal is weighed against the cost of an unnecessary one, and the regret of a conformal advisory is bounded in terms of coverage shortfall and the asymmetry of the loss. A replication-based evaluation of the ordinary least squares estimator was conducted to characterise the sampling behaviour that any calibration procedure must accommodate: 600 data sets of N = 500 were drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2). The estimator recovered the exposure coefficient with negligible bias (-0.00058) and empirical coverage of the nominal 95% intervals was 1.0 for both intercept and exposure, establishing a stable baseline against which shifted-calibration behaviour can be reasoned. Coverage alone is an insufficient criterion for advisory systems, and that calibration coverage and replanting-decision regret must be evaluated jointly.