Font Size: 

The use of innovative machine learning techniques has significantly transformedthe banking sector, including the credit origination process. The adoptionof more robust approaches has led to the creation of more accurate classificationmodels. However, this progress has been accompanied by a dilemma in high-levelbanking discussions: the lack of explainability and interpretability in black-box modelsinvolved in credit access decisions. This is particularly critical issue in banking,as customers and regulators expect to have a reasonable understanding of the variablesthat could improve or deteriorate your chances for credit access. In this study,we will discuss a methodological approach for a credit score for micro, small, andmedium-sized enterprises (MSMEs) at a commercial bank in Costa Rica. To addressthe interpretability challenge, we introduce both local and global interpretabilityperspectives, highlighting the GamiNet method as an enhanced example for neuralnetworks. Developed in recent years, GamiNet aims to maintain the robustness ofa generalized feedforward neural network approach with multiple additive subnetworks.Each subnetwork consists of several hidden layers, allowing us to capturemain effects and pairwise interactions. GamiNet is globally interpretable by design,and also demonstrates competitive accuracy compared to other more established machinelearning methods such as Random Forests.We also performvarious sensibilityanalysis on the model to assess its robustness.

Banking, Credit Score, Black Box, Neural Network, GamiNet, Random Forests