An end-to-end deep learning approach to credit scoring using CNN C XGBoost on transaction data

The Value of Open Banking Data for Application Credit Scoring: Case Study of a Norwegian Bank

“…There are several examples of deep learning in credit risk in recent years, both for consumer default prediction (Addo et al 2018;Dastile and Celik 2021;Gunnarsson et al 2021;Ha et al 2019;Hamori et al 2018;Hjelkrem et al 2022;Kvamme et al 2018;Shen et al 2021;Sirignano et al 2016;Wang et al 2018;Wu et al 2021) and bankruptcy prediction (Hosaka 2019;Jang et al 2021;Mai et al 2019;Shetty et al 2022;Smiti and Soui 2020;Stevenson et al 2021). We observe that most of these studies use a shallow learning approach; e.g., the deep learning algorithms are applied on conventional credit risk data sets where raw data are aggregated (typically by hand by experts) into explanatory variables.…”

“…This entails that the deep learning algorithms are applied directly on raw, unaggregated credit data, such as customer transactions and textual disclosures from loan applications, replacing handmade feature engineering with algorithmic creation of features using representation learning. Notable examples are Kvamme et al (2018), who successfully applied deep learning algorithms directly on daily balances from current accounts, Hjelkrem et al (2022) andAla'raj et al (2022), who applied deep learning algorithms on raw financial transaction data, while Stevenson, Stevenson et al (2021) and Mai et al (2019) successfully applied deep learning algorithms on raw text.…”

“…The CNN model used in this paper was developed by Hjelkrem et al (2022) on a data set containing approximately 150,000 observations. The network architecture is presented in Figure 1.…”

“…were decided on the basis of model performance on the validation data. See Hjelkrem et al (2022) for further details on model architecture, hyperparameters, and model training.…”