Predicting creditworthiness is an important task in the banking industry, as it allows banks to make informed lending decisions and manage risk. In this paper, we investigate the performance of two different deep learning credit scoring models developed on the textual descriptions of customer transactions available from open banking APIs. The first model is a deep learning model trained from scratch, while the second model uses transfer learning with a multilingual BERT model. We evaluate the predictive performance of these models using the area under the receiver operating characteristic curve (AUC) and Brier score. We find that a deep learning model trained from scratch outperforms a BERT transformer model finetuned on the same data. Furthermore, we find that SHAP can be used to explain such models both on a global level and for explaining rejections of actual applications.

Banks generally use credit scoring models to assess the creditworthiness of customers when they apply for loans or credit. These models perform significantly worse when used on potential new customers than existing customers, due to the lack of financial behavioral data for new bank customers. Access to such data could therefore increase banks’ profitability when recruiting new customers. If allowed by the customer, Open Banking APIs can provide access to balances and transactions from the past 90 days before the score date. In this study, we compare the performance of conventional application credit scoring models currently in use by a Norwegian bank with a deep learning model trained solely on transaction data available through Open Banking APIs. We evaluate the performance in terms of the AUC and Brier score and find that the models based on Open Banking data alone are surprisingly effective in predicting default compared to the conventional credit scoring models. Furthermore, an ensemble model trained on both traditional credit scoring data and features extracted from the deep learning model further outperforms the conventional application credit scoring model for new customers and narrows the performance gap between application credit scoring models for existing and new customers. Therefore, we argue that banks can increase their profitability by utilizing data available through Open Banking APIs when recruiting new customers.

The authors put forward an explainable machine learning model predicting credit default using a real-world data set provided by a Norwegian bank.

The authors find that machine learning methods can generate satisfactorily performing credit score models based on data from the 90-days prior to the score date

This paper examines the stability and accuracy of credit ratings from two Norwegian savings and loans banks, labeled Bank A and Bank B. Credit Rating Agencies (CRAs) often claim that ratings are relative rankings of firms and largely independent of the business cycle. We find that the intensity of banks’ rating changes – both upgrades and downgrades – vary over time depending on the business cycle. This is inconsistent with characterizing their methodology as through-the-cycle. We further find that the volatility and accuracy of Bank B – the bank with more customers exposed to the petroleum industry – seems to be higher than that of Bank A. The accuracy of Bank B’s ratings also appears to be less affected by economic slowdowns than those of Bank A. Whereas Bank A’s accuracy drops significantly following the oil price shock in 2014–2015, the accuracy of Bank B remains more stable.