GlucoFM-Bench

Time-series foundation models promise strong zero-shot forecasting across domains. Blood glucose is a demanding test of that promise: the signal is high-frequency, highly non-stationary, and driven by unobserved exogenous events such as meals, insulin, and exercise.

GlucoFM-Bench is a systematic benchmark evaluating how well modern time-series foundation models and LLM-based architectures forecast blood glucose, compared against established deep learning baselines.

Scope

• Evaluated time-series foundation models (Chronos, TimesFM), LLM-for-time-series approaches (TimeLLM), and conventional architectures (LSTM, Transformer) under a common protocol.
• Tested across uni- and multimodal data from both medical-grade and consumer wearables.
• Standardized preprocessing, forecast horizons, and evaluation metrics so that differences reflect the models rather than the experimental setup.

The benchmark provides a common reference point for whether general-purpose foundation models transfer to physiological forecasting, and where they still fall short of task-specific models (Lu et al., 2026).