Evolutionary-based ensemble feature selection technique for dynamic application-specific credit risk optimization in FinTech lending

This study introduces EFSGA, an evolutionary-based ensemble learning and feature selection technique inspired by the genetic algorithm, tailored as an optimized application-specific credit classifier for dynamic default prediction in FinTech lending. Our approach addresses existing gaps in metaheuristic applications for credit risk optimization by (i) hybridizing metaheuristics with machine learning to accommodate the dynamic nature of time-evolving systems and uncertainty, (ii) leveraging distributed and parallel computing for real-time solutions in complex risk decision processes, and (iii) enhancing applicability to unbalanced learning scenarios. The proposed model utilizes a heterogeneous ensemble of machine learning algorithms, incorporating a genetic algorithm to simultaneously optimize model hyperparameters and classification thresholds based on decision-maker objectives over time. This approach substantially improves out-of-sample model performance, providing valuable insights for timely post-loan risk management. The feature selection technique contributes to a balanced trade-off between model performance and interpretability—a pivotal consideration in metaheuristic-based models. Results obtained from the EFSGA model applied to a dataset spanning 2007 to 2014 unveiled an average improvement of 23% in application-specific evaluation metrics compared to conventional heterogeneous ensemble techniques across diverse risk-taking scenarios. Noteworthy is the proposed dynamic framework, featuring a tunable class-weighted fitness function, demonstrating significant superiority in delivering real-time solutions adaptable to evolving decision processes. We validate the EFSGA classification model against established credit evaluation models.