Censoring-Aware In-Context Learning for Generalized Supplier Lead Time Estimation in Supply Chain Planning

Supplier lead time forecasting is a critical component of supply chain management, yet it faces two persistent challenges: the need for models that generalize across diverse industrial settings without costly retraining, and the difficulty of incorporating 'open orders'—orders that have been placed but not yet received—into predictive models. Standard regression approaches often discard these open orders, while traditional survival models require complex, task-specific modeling. This paper introduces LT-ICL to address these gaps by treating lead time estimation as a right-censored probabilistic forecasting problem.

LT-ICL combines a transformer encoder with a conditional normalizing-flow head to produce flexible predictive distributions. The model is pretrained on synthetic, right-censored lead time tasks, allowing it to learn generalizable patterns of lead time behavior. By using an in-context learning framework, LT-ICL can adapt to new, proprietary industrial datasets at inference time without requiring any parameter updates. The authors also provide a theoretical decomposition of forecasting risk, identifying prior misspecification and amortized approximation errors as the two primary drivers of model performance.

The authors evaluated LT-ICL on 24 proprietary datasets spanning seven different industries. The model achieved the lowest point-forecasting error on 15 datasets and the lowest probabilistic forecasting error on 14 datasets, outperforming classical machine learning, deep learning, and other in-context learning baselines. These results demonstrate that amortized pretraining on synthetic data is a viable strategy for creating reusable, high-performance forecasting systems in industrial planning.

For practitioners, LT-ICL shifts the burden of model development from repeated, per-dataset retraining to a one-time upfront pretraining cost. This significantly reduces deployment overhead when scaling forecasting capabilities across different suppliers or product lines. Furthermore, by natively handling right-censored data, the model extracts more value from existing procurement data, leading to more reliable inventory and material requirements planning.