Forecasting COVID-19 cases based on a parameter-varying stochastic SIR model

João P Hespanha; Raphael Chinchilla; Ramon R Costa; Murat K Erdal; Guosong Yang

doi:10.1016/j.arcontrol.2021.03.008

Forecasting COVID-19 cases based on a parameter-varying stochastic SIR model

Annu Rev Control. 2021:51:460-476. doi: 10.1016/j.arcontrol.2021.03.008. Epub 2021 Apr 8.

Authors

João P Hespanha¹, Raphael Chinchilla¹, Ramon R Costa², Murat K Erdal¹, Guosong Yang¹

Affiliations

¹ University of California, Santa Barbara, USA.
² Federal University of Rio de Janeiro, Brazil.

Abstract

We address the prediction of the number of new cases and deaths for the coronavirus disease 2019 (COVID-19) over a future horizon from historical data (forecasting). We use a model-based approach based on a stochastic Susceptible-Infections-Removed (SIR) model with time-varying parameters, which captures the evolution of the disease dynamics in response to changes in social behavior, non-pharmaceutical interventions, and testing rates. We show that, in the presence of asymptomatic cases, such model includes internal parameters and states that cannot be uniquely identified solely on the basis of measurements of new cases and deaths, but this does not preclude the construction of reliable forecasts for future values of these measurements. Such forecasts and associated confidence intervals can be computed using an iterative algorithm based on nonlinear optimization solvers, without the need for Monte Carlo sampling. Our results have been validated on an extensive COVID-19 dataset covering the period from March through December 2020 on 144 regions around the globe.

Keywords: Epidemic models; Estimation; System identification.