Federated Learning Architectures for Credit Risk Assessment: A Comparative Analysis of Vertical, Horizontal, and Transfer Learning Approaches

The rapid evolution of financial services, spurred by the ascent of FinTech, has significantly transformed how we assess credit risk. Traditional centralized approaches to credit risk modelling, now struggle with issues surrounding data privacy, security, and the ability to scale. Against this backdrop, federated learning (FL) has come to the fore as a revolutionary strategy, enabling collaborative model crafting while ensuring data remains decentralized and private. This paper delves deeply into a comparative assessment of three key federated learning architectures—vertical federated learning (VFL), horizontal federated learning (HFL), and federated transfer learning (FTL), in the context of credit risk assessment in FinTech. We assess the structural aspects of each federated learning architecture, discussing their strengths, limitations, and key performance indicators within the credit risk context. Furthermore, the paper identifies key research gaps, open problems, and future research directions, emphasizing the need for advanced privacy-preserving techniques, handling data heterogeneity, ensuring fairness and explainability, improving scalability and efficiency, integrating with emerging technologies, and developing appropriate governance and regulatory frameworks. The comparative analysis presented in this paper provides valuable insights for researchers, practitioners, and policymakers working at the intersection of federated learning and credit risk assessment in FinTech. By understanding the strengths and limitations of different federated learning architectures, stakeholders can make informed decisions when designing and implementing collaborative credit risk modelling solutions. Lastly, the paper proposes an enhanced Adaptive Hierarchical Hybrid Federated Learning (HHFL) architecture, tailored to address these challenges and advance the field of credit risk assessment through federated learning.