Benchmarking a time-series foundation model (TimeGPT) for real-world forecasting applications

Accurate forecasting of hospital demand is essential for operational resilience, yet traditional statistical and machine learning approaches often require extensive feature engineering and tuning, limiting adoption in resource-constrained environments. Foundation models for time-series forecasting offer the potential for robust, zero-shot performance across domains. This study evaluates the feasibility of TimeGPT, a general-purpose time-series foundation model, for forecasting daily Emergency Department (ED) arrivals. We benchmarked TimeGPT against Seasonal Autoregressive Integrated Moving Average (SARIMAX), Prophet, and XGBoost under univariate and multivariate configurations. The experimental design simulated operational constraints by limiting the training window to 30 days and using a rolling forecast over a 60-day holdout period. Forecast accuracy was assessed using root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and directional accuracy. TimeGPT consistently ranked among the top-performing models. In the univariate setting, it achieved a MAPE of 7.7% and directional accuracy of 75%, comparable to or exceeding traditional models with extensive feature engineering. TimeGPT required no model-specific tuning and maintained accuracy without exogenous features such as weather or calendar variables. SARIMAX achieved the best results in the temporal-plus-weather configuration (MAPE 7.0%, RMSE 31.0) but required substantially more configuration. TimeGPT recorded zero large-error days (>30% deviation), while SARIMAX had 5 such days, underscoring the trade-off between accuracy and robustness. This benchmark demonstrates that foundation models can deliver accurate, reliable forecasts in healthcare operations with minimal data preparation. TimeGPT’s zero-shot capability highlights its potential as a scalable solution for diverse operational forecasting challenges.