Network Auto-Logistic Regression Model for Longitudinal Trade Credit Insurance Claim Data

This is the authors' abstract. We don't add key points for this paper.

Trade credit insurance (TCI) protects sellers against buyer defaults in open-account transactions. Accurate, scalable predictions of claim occurrence at the buyer-seller connection level, accounting for serial dependence and high-order network contagion, are essential for ratemaking. To this end, we introduce an extended network auto-logistic regression (NAR) model tailored to this setting. The model augments logistic regression with operators built from the directed buyer-seller graph to encode independent covariates, pairwise adjacency, two-star and two-step topologies, and cross-temporal effects, thereby capturing higher order network and longitudinal dependence among connections. Parameters are estimated via maximum pseudo-likelihood (MPLE), which reduces to standard logistic regression fitting on an expanded feature set, yielding a convex objective with a unique solution, fast computation, and straightforward interpretation. Applied to a large TCI dataset, the extended NAR delivers superior out-of-sample predictive power relative to both network and nonnetwork benchmarks with minimal computational burden, while providing interpretable risk attributions linked to network structure.