Prospective comparison of econometric, machine learning, and foundation models for forecasting emergency department boarding patients.

Emergency department (ED) boarding drives overcrowding, worsens outcomes, and strains hospital operations. Accurate short-term forecasts of boarding volumes can enable proactive resource management. In this prospective study, we compared six forecasting approaches across econometric, machine-learning, and foundation-model paradigms to predict ED boarding volumes up to four days ahead (T + 1-T + 4). Using data from two UC San Diego Health EDs (October 2022-October 2024), models incorporated prior boarding volumes, scheduled surgeries, hospital census, and expert-selected covariates. We evaluated vector autoregression (VAR), extreme gradient boosting (XGBoost), and Google TimesFM, including hybrids (VAR+XGBoost; TimesFM+XReg), against a two-week moving-average baseline. During a four-month validation period, the VAR+XGBoost hybrid achieved the lowest root mean square error-reducing forecast error by 16-42% at La Jolla and 5-19% at Hillcrest-while showing only slight gains over VAR (0-3%). VAR alone performed robustly and outperformed baseline at both sites. The VAR model has been deployed within the health system's Mission Control to guide proactive interventions such as discharge acceleration and surgical rescheduling. These findings underscore the enduring value of econometric models and demonstrate how forecast-driven decision support can enhance emergency care coordination and system responsiveness.