Improvement of Credit Scoring by LSTM Autoencoder Model

One of the most important causes of the financial crisis of 2007–2008 is the risk management systems. Credit scoring is an essential part of these risk management systems in financial institutions. While the effectiveness of feature extraction by artificial intelligence on credit scoring is well known, the effect of historical repayment data needs to be studied. The performance of machine learning models depends on the features and representation of data. The features extracted by deep learning models in an end-to-end manner have more information than hand engineered features. In this paper, we extract features from customer’s previous loan repayments in two ways: manual and by LSTM autoencoder, and their effects evaluated on the credit scoring problem. Using proposed methods to feature extraction, shows improvement in the performance metrics of the models.