Analysis of the probability of default in peer-to-peer lending. Application of different classification techniques.

Abstract

In this thesis, peer-to-peer lending is explored and analyzed with the objective of fitting a model to accurately predict if borrowers default on their loans or not. The foundation for the thesis is a dataset from LendingClub, a peer-to-peer lending platform based in San Francisco, USA. Detailed information of borrowers’ financial history, personal characteristics and the specifics of each loan is used to predict the probability of default for the various loans in the portfolio. Methods used include elastic net regularization of logistic regression, boosting of decision trees, and bagging with random forests. The results are compared using accuracy metrics and a profitability measure, before a final model selection is carried out.