Multi-Armed Bandit to optimize the pricing strategy for consumer loans

Abstract: This thesis explores the possibility of framing the problem of setting prices for consumer loans on loan comparison sites as a Multi-Armed Bandit Problem. The problem is solved by creating a Multi-Armed Bandit environment based on SEB:s expert knowledge of the problem. Different Multi-Armed Bandit algorithms are then compared in a stationary environment after which the best performing algorithm is modified to handle a non-stationary environment. We found that the Sliding-Window Thompson Sampling is the best choice of algorithm for the problem. Furthermore, we show that this method is not sensitive to the assumptions made when generating the non-stationary environment, thus making it a promising method for real-world application.