Theta Autoregressive Neural Network: A Hybrid Time Series Model for Pandemic Forecasting

ABSTRACT

Forecasting time series present a perpetual topic of research in statistical machine learning for the last five decades. Due to the unprecedented outbreak of the novel coronavirus (COVID-19), forecasting the COVID-19 pandemic became a key research interest for both epidemiologists and statisticians. These future predictions are useful for the effective allocation of health care resources, stockpiling, and help in strategic planning for clinicians, government authorities, and public-health policymakers. This paper develops an effective forecasting model that can generate real-time short-term (ten days) and long-term (fifty days) out-of-sample forecasts of COVID-19 outbreaks for eight profoundly affected countries, namely the United States of America, Brazil, India, Russia, South Africa, Mexico, Spain, and Iran. A novel hybrid approach based on the Theta method and Autoregressive neural network (ARNN) model, named Theta-ARNN (TARNN) model, is proposed. The proposed method outperforms previously available single and hybrid forecasting models for COVID-19 predictions in most data sets. The ergodicity and asymptotic stationarity of the TARNN model are also studied.