A Relational Data Matching Model for Enhancing Individual Loss Experience : An Example from Crop Insurance

The focus of this paper is on predictive analytics regarding data scarcity and credibility, which are major difficulties facing the agricultural insurance sector often due to limited loss experience data for those infrequent, but, extreme weather events. A new relational data-matching model is presented to predict individual farmer yields in the absence of farm-level data. The relational model defines a similarity measure based on an Euclidean distance metric that considers weather information, farm size,county size and the coefficient of variation of yield, to search for the most "similar" region in a different country to borrow individual loss experience data that is otherwise not available. Detailed farm-level and county-level corn yield data in Canada and the U.S. are used to empirically evaluate the proposed relational model. Compared to the benchmark model, the relational model developed in this paper has lower prediction error with smaller variation, and is also able to recover the actual premium rate more accurately. While the example in this paper uses historical data from Canada and the U.S. to illustrate the approach, the model may be extended to other countries, including developing countries, where farm-level data may not be available. This provides a new approach for insurers, reinsurers, and governments to enhance risk modelling,pricing, and the development of new insurance programs