A Financial-Logistics Graph Framework for Dynamic Mortgage Default Prediction Using Recurrent Hazard Modeling

Aleksey Evgenievich Shulga, Sergey Evgenievich Barykin, Daria Dinets

Keywords: Financial logistics, mortgage default prediction, survival analysis, hazard modeling, recurrent neural networks, graph-based modeling, ablation study, time-dependent AUC, credit risk analytics

Abstract: The paper proposes a financial and logistical graph model for dynamic forecasting of mortgage default, in which the borrower is formalised as a time-dependent system of cash flows. Unlike traditional scoring approaches based on static characteristics, default is interpreted as the result of the degradation of the flow structure and the depletion of the buffer stability of the borrower's financial graph. A graph-recurrent hazard architecture has been developed that represents the borrower as a four-node directed graph - structural state (S), flow dynamics (F), buffer capacity (B), and disruption indicators (D) — with edges encoding economically motivated dependencies between components. A Graph Convolutional Network (GCN) performs message passing over this internal financial graph at each timestep, producing graph-level embeddings that are then processed by a GRU-based recurrent network to estimate temporal default intensity. A formal rule for matching empirical features of Fannie Mae panel data with components of the financial-logistical framework is proposed, which ensures the methodological transparency of group ablation. Empirical testing was conducted on panel data of mortgage loans using ROC-AUC, PR-AUC, precision@k, Brier score, and time-dependent AUC metrics. All results are reported with confidence intervals from multiple training runs with different random initializations. The proposed model demonstrates consistent superiority over logistic regression, gradient boosting, random forest, static neural network, and Cox's model. Ablation analysis confirms the dominant role of flow dynamics compared to static borrower characteristics. Structural ablation comparing the proposed GNN+GRU model against a vanilla GRU baseline confirms that the graph component provides measurable improvement in predictive performance, validating the computational value of the financial-logistics graph structure beyond conceptual framing. The results obtained indicate that mortgage default should be considered as systemic instability of a dynamic financial structure, which opens up opportunities for the development of more interpretable and adaptive credit risk management models.

Submission Number: 67