Fitting dynamic models to epidemic outbreaks with quantified uncertainty: A primer for parameter uncertainty, identifiability, and forecasts

Chowell, Gerardo

We are frequently interested in calibrating a model not only to understand and characterize the current state of the system, but also to aim to predict its behavior in the near or long terms. The particular time horizon of forecast depends on the purpose of the forecast. For instance, a long-term forecast (e.g., several years) could be useful to make strategic decisions regarding the construction of facilities to respond to natural disasters such as epidemics and hurricanes whereas a short-term forecast (e.g., days to weeks) are useful to plan for scheduling resources (e.g., number of face masks, hospital beds). However, it is important to keep in mind that forecasts are often inaccurate as these are mostly based on the current values and uncertainty of the parameters of the system, which are likely to change over time. Moreover, the further out the forecast is made, the more wrong it is expected to be. A properly calibrated model to data can be used to generate short-term or long-term forecasts of the system. Generating a forecast based on the model uncertainty given by f ðt; b