Association of Automated Feature Engineering with Default Prediction in Digital Banking Records

Abstract

The transition from traditional retail banking to digital banking ecosystems has resulted in an exponential increase in the volume, velocity, and variety of customer data. Accurately assessing credit risk and predicting default probabilities in this environment requires advanced analytical techniques capable of capturing complex behavioral patterns. While machine learning algorithms have significantly improved predictive performance in credit scoring, the efficacy of these models fundamentally relies on the quality of feature engineering. This study presents a comprehensive benchmarking analysis of automated feature engineering techniques applied to digital banking records for default prediction. By systematically comparing traditional manual feature engineering with automated methods utilizing deep feature synthesis, this research evaluates predictive accuracy, interpretability, and computational efficiency across a spectrum of machine learning classifiers. The empirical results demonstrate that automated feature engineering consistently enhances the discriminatory power of credit scoring models by uncovering latent temporal and relational data structures that manual processes often overlook. Furthermore, the generated behavioral features provide novel insights into customer risk profiles, albeit at the cost of increased computational complexity. This paper provides a robust framework for integrating automated feature generation into financial risk management, offering actionable insights for institutional practitioners and establishing a foundation for future academic inquiry in automated machine learning applications within the financial sector.