Researchers develop specialized LLM framework that forecasts supply chain disruptions more accurately than GPT-5 by training on real disruption outcomes

New end-to-end framework trains large language models to produce calibrated probabilistic forecasts specifically for supply chain disruptions

Custom model outperforms strong baselines including GPT-5 on accuracy, calibration, and precision metrics

Training approach enables models to develop structured and reliable probabilistic reasoning without explicit prompting

Framework addresses the challenge of predicting rare, high-impact events from noisy and unstructured data

Researchers released evaluation dataset on Hugging Face to support transparency and reproducibility