Memory effect in time-window epidemic curve forecasting using Approximate Bayesian Computation

Fitting compartmental models to epidemiological data aiming to produce reasonable forecasts can become a very complex task, especially when the data assume a behavior difficult to be attained by models with constant parameters. A common alternative is to build models with time-dependent parameters, which does not necessarily simplify the fitting process, but can make the model more descriptive. In this work we propose to adopt a simple SEIRD model with constant parameters, but dividing the epidemiological data into different time-windows, in which it is assumed that the data can be piecewise fitted, as an alternative way of adopting time-dependent parameters. Using Approximate Bayesian Computation , posterior distributions of parameters obtained in previous windows are used as prior distributions of corresponding parameters in subsequent windows. We show that taking advantage of this information does improve the predictive capacity of the model, when compared to the strategy in which noninformative priors are adopted for each window. Finally, we assess the combination of time-windows with different lengths, seeking for more accurate forecasts.