Evaluating forecasting models for health service demand during the COVID-19 pandemic

We combine daily internet search data and monthly information on medical expenditures for anti-depressants to test two distinct hypotheses in eight Australian states, covering the period from 2020 to 2022. First, whether using daily search data can help predict future demand for health services; and second, whether nowcasting and machine learning models yield better predictions vis-à-vis autoregressive forecast models. Our aim is to assess the use of such data and techniques to improve the planning and possible deployment of health resources across space. We find that search data contain information that is valuable for predicting the demand for health services in the short-run, and that machine learning models yield predictions with lower mean square error. Both results support the use of daily search data and machine learning tools to enhance the provision of health services across locales.