Temporally Aware Generative Adversarial and Graph Convolutional Framework for Horizon-Specific Loan Default Prediction
Year: 2025
Type: article
Abstract: The detection of mortgage defaults, with respect to time, is hampered by severe class imbalance, changing risk profiles throughout forecasting periods, and the exclusion of borrower-network effects in conventional models. Traditional machine learning models fail to fully address these complexities, frequently leading to unreliable default predictions and ineffective risk management. We address these gaps with a two-part framework that (1)uses a time-conditioned conditional generative adversarial network to generate horizon-specific minority examples, and (2) adds borrower similarity to a graph convolutional network using metropolitan-area embedding and a postal-code proximity graph. We apply the pipeline to a curated subset of Freddie Mac single-family loan records and compare against oversampling and flat-classifier baselines. Incorporating borrower network information yields the best performance; the graph model consistently elevates default detection and shows that network features (for example, lender identity and geographic clusters) are dominant, with top risk drivers varying over time. To help with reproducibility, we make available full training settings, architectures, and evaluation processes. Limitations include the U.S. fixed-rate mortgage scope and historical data cutoffs.
Authors: Aditya Kariwal, Rishit Singh
Institution: Indian Institute of Technology Kharagpur
Subfield: Accounting
Domain: Social Sciences
SDG: Reduced inequalities
Open Access status: bronze
OpenAlex ID: W4415735660