Benchmarking Time-Series Foundation Models and Multimodal Dietary Context for CGM Forecasting

Zero-shot time-series foundation models fail to consistently outperform task-specific baselines like Elastic Net or PatchTST on continuous glucose monitoring (CGM). In a benchmark spanning eight datasets across diabetic and non-diabetic cohorts, lightweight fine-tuning of Chronos-Bolt reduced forecasting RMSE by up to 18.4%. Incorporating multimodal dietary context—food images and macronutrient logs—via residual fusion reduced postprandial forecasting error by an additional 15%, showing that biological time-series forecasting requires both domain adaptation and external signals.