Evaluation of mechanistic and statistical methods in forecasting influenza-like illness

A variety of mechanistic and statistical methods to forecast seasonal influenza have been proposed and are in use; however, the effects of various data issues and design choices (statistical versus mechanistic methods, for example) on the accuracy of these approaches have not been thoroughly assessed. Here, we compare the accuracy of three forecasting approaches—a mechanistic method, a weighted average of two statistical methods and a super-ensemble of eight statistical and mechanistic models—in predicting seven outbreak characteristics of seasonal influenza during the 2016–2017 season at the national and 10 regional levels in the USA. For each of these approaches, we report the effects of real time under- and over-reporting in surveillance systems, use of non-surveillance proxies of influenza activity and manual override of model predictions on forecast quality. Our results suggest that a meta-ensemble of statistical and mechanistic methods has better overall accuracy than the individual methods. Supplementing surveillance data with proxy estimates generally improves the quality of forecasts and transient reporting errors degrade the performance of all three approaches considerably. The improvement in quality from ad hoc and post-forecast changes suggests that domain experts continue to possess information that is not being sufficiently captured by current forecasting approaches.