On the Application of Neural Networks in Credit Scoring: A Case-Study Analysis

This study investigates the efficacy of neural network models for credit scoring using a substantial dataset from Kaggle, consisting of 80,000 training records and 20,000 testing records. We explore various data preparation techniques, including multiple imputation methods and normalization, to address missing values and scale feature magnitudes. Employing a Sequential neural network architecture with three hidden layers, we demonstrate the model's predictive accuracy and stability through cross-validation, achieving a Mean Absolute Percentage Error (MAPE) of 4.379%. Despite the promising results, we also discuss the complexities associated with selecting optimal model parameters and architectures. Future research directions are proposed, emphasizing the potential for neural network models to efficiently handle missing data and improve credit scoring processes.