Machine Learning for Optimal Index Insurance

Abstract:

Agricultural index insurance contracts have traditionally been developed using a two step process. First, losses are estimated based on crop yield, weather, and/or satellite data. Second, those loss estimates are used to design an insurance contract. Machine learning methods make it possible to combine both steps to generate the best possible insurance contract based on the data available. To achieve this result, I use expected utility theory to define the objective (loss) function of a simple neural network. The resulting algorithm uses satellite data coupled with data on farmer incomes and a fixed rate of profit for the insurer to identify the insurance contract that maximizes farmer welfare. The result is an insurance contract that is better for farmers and greatly simplifies index and contract design. The method I propose also makes it relatively simple to quickly test whether a data source has the potential to serve as an effective insurance index. Finally, the method makes it possible to develop different contracts for varying levels of risk aversion, or to generate optimal contracts based on theories from behavioral economics such as loss aversion, rank dependence, or cumulative prospect theory.