Associated factors and spatiotemporal patterns of agricultural insurance risk control using K-means clustering and anomaly detection

Abstract

Agricultural insurance risk control requires comprehensive information about persistent exposure, geographically correlated losses, and exceptional portfolio conditions to ensure financial stability. This study combines United States Department of Agriculture insurance records with National Oceanic and Atmospheric Administration climate data for 48 contiguous states during 2010–2024. The 720 state-year observations are analyzed using standardized K-means clustering, Isolation Forest screening, and spatial randomization tests; fixed-effects models use 672 observations after constructing annual lags. Silhouette selection favors two distinct profiles. The elevated-loss profile contains 19.7% of observations and 15.9% of insured liabilities but accounts for 44.4% of total indemnities. Annual loss-cost ratios exhibit positive spatial association in 14 of 15 years after multiple-testing adjustment. Among 36 observations selected under a 5% review budget, 16 are also selected by Local Outlier Factor. Regression estimates associate greater dry-side moisture pressure and higher April–September temperature anomalies with higher insured losses, while wet-side and financial associations are more sensitive to specification. Cluster agreement remains high under alternative linear scaling but weakens substantially after logarithmic compression. These results support geographically coordinated monitoring and review of exceptional observations while showing why cluster labels, detector agreement, and aggregate financial ratios cannot independently establish management effectiveness or misconduct. Ultimately, the contribution is a reproducible framework linking portfolio classification to spatial diagnostics and explicitly qualified factor analysis, providing actionable insights for policymakers and insurance providers to mitigate climate-related financial risks effectively.