Double Machine Learning for Insurance Price Optimization

Sammanfattning: This thesis examines how recent advances in debaised machine learning can be used for estimating price elasticities of demand within the automotive insurance field. Traditional methods such as generalized linear model (GLM) to estimate demand has no way of ensuring there are no biases in the underlying data selection, especially when the confounding variables are many. These approaches instead rely on the user’s experience to remove biases in the data. Advances, in the crossing fields between economics and machine learning have however found new approaches to debias datasets automatically through the double machine learning approach (DML). Using a large data set of insurance offers and sales from a Swedish insurance company, the double machine learning approach first described by Chernozhukov et al. (2016) is used to estimate the price elasticity of demand for individual customers. The price elasticities are then grouped on variables of importance and combined with the loss ratio of the segment in order to optimize the existing insurance tariff. In terms of model performance, the importance of the first stage classifier and regressor proved to be important for the final results. In alignment with expectations of the results, higher premium cars such as BMW and Mercedes proved to be more price sensitive. However, these brands also had higher loss ratios which resulted in a lower potential for lowering prices. Aligning the price elasticity with the loss ratios and the company’s strategy was found to be an important aspect. On average, the automotive insurance industry was shown to be price sensitive with few segments of inelastic characteristics.