Forecasting Supply Chain Disruptions with Foresight Learning

A team of researchers including Benjamin Turtel, Paul Wilczewski, and Kris Skotheim has published a paper introducing 'Foresight Learning,' a novel framework designed to tackle the critical challenge of predicting rare, high-impact supply chain disruptions. The core innovation is an end-to-end method that trains large language models (LLMs) to generate calibrated probabilistic forecasts. It uses actual, realized disruption outcomes as the supervisory signal, teaching the model to reason reliably from noisy, unstructured data—a scenario where general-purpose models like GPT-5 typically struggle without significant task-specific adaptation.

In their evaluations, the resulting Foresight Learning model demonstrated substantial performance gains. It outperformed strong baselines, explicitly noted to include GPT-5, across key metrics: accuracy, calibration (the reliability of its probability estimates), and precision. The research also shows that this training process induces more structured and reliable probabilistic reasoning within the model without the need for complex, explicit prompting techniques. This suggests a generalizable pathway for creating specialized, decision-ready forecasting tools for various domains. To promote transparency and further development, the researchers have open-sourced the evaluation dataset used in their study.