A0311
Title: PDx: Adaptive credit risk forecasting model in digital lending using machine learning operations
Authors: Sayantan Banerjee - Indian Institute of Management Indore (India) [presenting]
Abstract: The aim is to present PDx, an adaptive, MLOps-driven system for forecasting credit risk using probability of default (PD) modeling in digital lending. Traditional PD models focus on accuracy at the development stage using complex ML algorithms, but often fail to adapt to evolving borrower behavior, leading to static models that degrade in production. Many lenders also struggle to deploy and maintain ML models effectively. PDx addresses these issues with a dynamic, end-to-end model lifecycle framework that includes continuous monitoring, automated retraining, and validation through a robust MLOps pipeline. A key innovation is a champion-challenger architecture, enabling regular updates and recalibration using recent data, ensuring resilience to data drift and shifting credit patterns. Empirical results show that ensemble tree models outperform others in default classification but require frequent updates to maintain performance. In contrast, logistic regression and neural networks exhibit faster degradation. By mitigating model decay and value erosion, PDx is especially effective for short-term, small-ticket digital loans where borrower behavior shifts quickly. PDx is validated using peer-to-peer, business, and auto loan datasets, demonstrating its scalability and adaptability for modern credit risk modeling.
