Frequency-Enhanced Retrieval-Augmented Wide Stacking for Supply-Chain Credit Risk Prediction
Abstract
Credit risk prediction in supply-chain finance is hard because the risk of a core enterprise is shaped not only by its own financials but also by the long-cycle operating conditions of its upstream suppliers. Conventional credit scoring models treat the task as static tabular classification, whereas time-series models tend to overlook the severe label imbalance and the heterogeneous evidence sources that characterize financial risk data. We propose FDRAG-WideStacking, a frequency-enhanced, dynamic, retrieval-augmented wide-stacking framework that jointly exploits four complementary evidence channels: static financial profiles, robust frequency-domain representations of upstream five-year time chains, risk-aware retrieval evidence from historical time clusters, and strong tabular expert scores. Unlike compact decision fusion, which compresses branch outputs before the final decision, wide stacking preserves raw discriminative features together with interpretable branch-level probabilities, a property we show is better suited to sparse positive labels. We further provide Proposed-FDRAG-Fusion, a validation-weighted extension for ranking-oriented screening. On 2,384 supply-chain finance samples containing only 74 risk cases, FDRAG-WideStacking attains the best F1 among all baselines and ablations, while the robust extension yields higher ROC-AUC and Average Precision with lower F1 variance.