AI Improves Supplier Lead Time Forecasting with Censoring-Aware Learning.

A new model, LeadTime-ICL, has been developed to improve supplier lead time forecasting, a critical component of supply chain planning. This model addresses the common challenge of right-censored data in industrial datasets, where some orders have not yet arrived when forecasts are needed, a limitation that standard regression models often discard. LeadTime-ICL combines a transformer backbone with a conditional normalizing-flow head to generate full predictive distributions over lead times, rather than just point estimates. The model is pretrained on synthetic right-censored tasks, enabling it to adapt to new industrial datasets in-context without requiring task-specific parameter updates. Evaluated on 24 proprietary supply chain datasets across seven industries, LeadTime-ICL achieved the lowest point-forecasting error on 15 datasets and the lowest probabilistic forecasting error on 14, demonstrating superior average rank. This research validates right-censored probabilistic forecasting as a practical approach and highlights the effectiveness of pretrained in-context models for accurate, low-adaptation-cost forecasting in industrial planning.