From Drought Risk Mapping To Parametric Insurance: A Machine Learning-Based Framework

Agricultural drought has become more frequent and severe under climate change, generating growing interest in parametric insurance products that rely on objective environmental indicators. This paper develops a flexible modeling framework for the design of index-based drought insurance using open-access climatic and satellite data. The methodology is based on the construction of synthetic indices that combine standardized precipitation indices (SPI and SPEI) with vegetation indicators (NDVI), selected through supervised learning methods such as Lasso regression, Random Forest, and Partial Least Squares. These composite indicators serve as the foundation for parametric indemnity functions that are specified as decreasing or increasing depending on the sign of their correlation with crop yields. The payout structure is calibrated with historical yield data and adjusted to minimize basis risk, while pure premiums are estimated through simulation-based methods including empirical resampling, bootstrapping, Monte Carlo, and kernel density estimation. The framework is illustrated with data from three Moroccan regions characterized by distinct agro-climatic conditions and dominant crops. The results show that the proposed design reduces basis risk significantly while maintaining transparency, interpretability, and reproducibility. Beyond this empirical application, the framework provides a generalizable contribution to the literature by integrating statistical learning with actuarial methods, offering a rigorous and transferable basis for the development of weather index insurance in data-scarce environments.