Risk scoring for crop insurance at enrollment: evidence and limits

Abstract

Crop insurance has to be priced and screened before the season’s main losses are known. This paper asks how far an insurer can get using only the information already available when a contract is enrolled: where the farm is, what crop and practice are insured, the chosen coverage level, and the local hail rate history. Using administrative records from 2006 to 2024, we build a practical underwriting score that separates the chance of a claim from the likely size of a loss, then test it on recent years and compare it with simpler rules and alternative models. The score ranks contracts better than regional averages, hail rate rules, or premium-based sorting, and the highest-ranked fifth of contracts contains about one third of the realized losses. Still, it misses much of the most severe loss risk. Coverage choice helps with prediction but also reflects farmer decisions, and thus the score should be read as a contract risk measure rather than a causal measure of hazard. These results suggest an enrollment-time information constraint for the specifications tested here, while leaving room for richer administrative and hazard-based extensions. Better tail prediction may require weather, farm history, and spatial information not used in the strict score.