Foundation Models for Sparse, Multi-Relational Risk Prediction in Global Supply Chains

Tabular and relational foundation models have demonstrated strong in-context learning on academic benchmarks, but their behavior on enterprise-scale structured data—marked by multi-relational schemas, extreme sparsity, and cold-start inference requirements—remains understudied. We evaluate two foundation model paradigms on global supply chain compliance risk prediction, a setting that stresses all three dimensions simultaneously. Our AutoFE+TFM pipeline automates temporal feature engineering across relational tables via Deep Feature Synthesis (DFS) and performs in-context learning without fine-tuning, discovering cross-table interaction features invisible to the hand-crafted production baseline. A relational foundation model (RFM), Griffin, preserves multi-table structure. On 3,700+ entities, DFS+TabICL improves P@90R from .662 to .665 and F1 from .745 to .750 when incorporating resolution tables, while matching baseline ROC-AUC (.804 vs. .807). On a disjoint-entity supplier split, a metadata-only inference mode retains .729 AUC and .622 P@90R for zero-history entities, exceeding the dataset positive rate (.540) and confirming actionable cold-start scoring. Our results show that foundation models match a gradient-boosted baseline on enterprise relational data while eliminating manual feature engineering and enabling cold-start inference—a capability absent from the current production pipeline.