Feasibility of very short-term forecast models for COVID-19 hospital-based surveillance

Abstract

Introduction: We evaluated the performance of Bayesian vector autoregressive (BVAR) and Holt's models to forecast the weekly COVID-19 reported cases in six units of a large hospital. Methods: Cases reported from epidemiologic weeks (EW) 12-37 were selected as the training period, and from EW 38-41 as the test period. Results: The models performed well in forecasting cases within one or two weeks following the end of the time-series, but forecasts for a more distant period were inaccurate. Conclusions: Both models offered reasonable performance in very short-term forecasts for confirmed cases of COVID-19.