Predicting COVID-19 using past pandemics as a guide: how reliable were mathematical models then, and how reliable will they be now?

During the earliest stages of a pandemic, mathematical models are a tool that can be implemented quickly. However, such models are based on meagre data and limited biological understanding. We evaluate the accuracy of various models from recent pandemics (SARS, MERS and the 2009 H1N1 outbreak) as a guide to whether we can trust the early model predictions for COVID-19. We show that early models can have good predictive power for a disease's first wave, but they are less predictive of the possibility of a second wave or its strength. The models with the highest accuracy tended to include stochasticity, and models developed for a particular geographic region are often applicable in other regions. It follows that mathematical models developed early in a pandemic can be useful for long-term predictions, at least during the first wave, and they should include stochastic variations, to represent unknown characteristics inherent in the earliest stages of all pandemics.