Tailoring time series models for forecasting coronavirus spread: Case studies of 187 countries

Abstract

When will the coronavirus end? Are the current precautionary measures effective? To answer these questions it is important to forecast regularly and accurately the spread of COVID-19 infections. Different time series forecasting models have been applied in the literature to tackle the pandemic situation. The current research efforts developed few of these models and validates its accuracy for selected countries. It becomes difficult to draw an objective comparison between the performance of these models at a global scale. This is because, the time series trend for the infection differs between the countries depending on the strategies adopted by the healthcare organizations to decrease the spread. Consequently, it is important to develop a tailored model for a country that allows healthcare organizations to better judge the effect of the undertaken precautionary measures, and provision more efficiently the needed resources to face this disease. This paper addresses this void. We develop and compare the performance of the time series models in the literature in terms of root mean squared error and mean absolute percentage error.

Keywords: COVID-19; Coronavirus; Epidemic transmission; Forecasting models; Machine learning models; Pandemic; Time series models.

Fig. 1

Long Term Short Memory Cell.

Fig. 2

Countries where ARIMA model achieves the lowset RMSE.

Fig. 3

Countries where ARIMA model achieves the lowset MAPE.

Fig. 4

Prediction of COVID-19 infections in Afghanistan for the Validation Dataset using the Time Series Models.

Fig. 5

Countries where Holt’s Linear Trend model achieves the lowest RMSE.

Fig. 6

Countries where Holt’s Linear Trend model achieves the lowest MAPE.

Fig. 7

Prediction of COVID-19 infections in Belarus for the Validation Dataset using the Time Series Models.

Fig. 8

Countries where S-curve Trend model achieves the lowest RMSE.

Fig. 9

Countries where S-curve Trend model achieves the lowest MAPE.

Fig. 10

Prediction of COVID-19 infections in Greece for the Validation Dataset using the Time Series Models.

Fig. 11

Countries where Damped Trend model achieves the lowest RMSE.

Fig. 12

Countries where Damped Trend model achieves the lowest MAPE.

Fig. 13

Prediction of COVID-19 infections in Australia for the Validation Dataset using the Time Series Models.

Fig. 14

Countries where Quadratic Trend model achieves the lowest RMSE.

Fig. 15

Countries where Quadratic Trend model achieves the lowest MAPE.

Fig. 16

Prediction of COVID-19 infections in Cameroon for the Validation Dataset using the Time Series Models.

Fig. 17

Countries where LSTM model achieves the lowest RMSE.

Fig. 18

Countries where LSTM model achieves the lowest MAPE.

Fig. 19

Prediction of Confirmed COVID-19 Cases in Bahamas for the Testing Dataset.

Fig. 20

Countries where Holt-winters’ Additive model achieves the lowest RMSE.

Fig. 21

Countries where Holt-winters’ Additive model achieves the lowest MAPE.

Fig. 22

Prediction of Confirmed COVID-19 Cases in San Marino for the Testing Dataset.

Fig. 23

Countries where SSM model achieves the lowest RMSE.

Fig. 24

Countries where SSM model achieves the lowest MAPE.

Fig. 25

Prediction of Confirmed COVID-19 Cases in Albania for the Testing Dataset.

Fig. 26

Countries where Sutte-ARIMA model achieves the lowest RMSE.

Fig. 27

Countries where Sutte-ARIMA model achieves the lowest MAPE.

Fig. 28

Prediction of Confirmed COVID-19 Cases in Nicaragua for the Testing Dataset.

Fig. 29

One-to-one Relationship between the Data Trend and its Accurate Time Series Model with Average MAPE.