Title: Credit scoring: Moving beyond the traditional approach with random forests and artificial neural networks
Authors: Samuel Stephan - Université Paris I Panthéon - Sorbonne (France) [presenting]
Matthieu Garcin - Leonard de Vinci Pole Universitaire (France)
Abstract: Traditional approaches, such as the logistic regression, are widespread in credit scoring because they are fast to implement, give quite accurate results, and allow for explainability. More advanced methods, such as ensembles (e.g. random forest, gradient boosting) or ANN (Artificial Neural Network), stay underused because of their complexity and their resulting lack of transparency. Practitioners are reluctant to use these techniques because they are known to have a high variance due to their large number of parameters which can make them unstable on new data. Indeed, these algorithms require knowledge on hyperparameters tuning in order to get workable results. We have compared three models: a logistic regression as a benchmark, a random forest, and an ANN. These algorithms have been fitted on a real credit dataset on individual loans sold from 2015 to 2018 by a French bank. The results show that random forest and ANN outperform the traditional logistic regression and would benefit to be used in financial applications.