EXTREME PROBABILITY FORECASTING ALGORITHM: CORONAVIRUSSPREAD IN SOUTH AFRICA

Abstract

Extreme temporal dynamics characterise the spread of coronavirus in South Africa. Thisresultsin uncertainties in estimating and forecasting,which can be handled usingaprobabilistic approach.We develop a combined forecast model by casting off a Markov-switching autoregressivecoupled with a truncated (or non-stationary) generalised extreme value (TGEVD)-based ensemble distribution. Exploratory data analysis revealed that the distribution of daily confirmed coronaviruscases is non-normal and that the series follows a fat-tailed distribution with significant regime shifts. Based on performance metrics such as root mean square error, mean absolute percentage error, and the continuous ranked probability score (CRPS), the novel hybrid model MS-AR-TGEV demonstrates the highest predictive accuracy for uncertainties in confirmed coronavirus cases.This study identifies the MS(2)-AR(1)-TGEV hybrid model as the most effective approach for forecasting uncertainties related to confirmed COVID-19 cases in South Africa. The result confirms the finding from other studies that hybrid models produce the most accurate performance in predicting time series data characterised by complex non-linear dynamics.The proposed model can be used by health sector personnel for planning purposes and for predicting any emerging epidemics that exhibit similar behaviour to the Coronavirus of 2019.