A hybrid of artificial neural network, exponential smoothing, and ARIMA models for COVID-19 time series forecasting

Abstract

The Auto Regressive Integrated Moving Average (ARIMA) model seems not to easily capture the nonlinear patterns exhibited by the 2019 novel coronavirus (COVID-19) in terms of daily confirmed cases. As a result, Artificial Neural Network (ANN) and Error, Trend, and Seasonality (ETS) modeling have been successfully applied to resolve problems with nonlinear estimation. Our research suggests that it would be ideal to use a single model of ETS or ARIMA for COVID-19 time series forecasting rather than a complicated Hybrid model that combines several models. We compare the forecasting performance of these models using real, worldwide, daily COVID-19 data for the period between January 22, 2020 till June 19, and June 20 till January 2, 2021 which marks two stages, each stage indicating the first and the second wave respectively. We discuss various forecasting approaches and the criteria for choosing the best forecasting technique. The best forecasting model selected was compared using the forecasting assessment criterion known as Mean Absolute Error (MAE). The empirical results show that the ETS and ARIMA models outperform the ANN and Hybrid models. The main finding from the ETS and ARIMA models analysis indicate that the magnitude of the increase in total confirmed cases over time is declining and the percentage change in the death rate is also on the decline. Our results shows that the chosen forecaste models are consistent during the first and second wave of of the pandemic. These forecasts are encouraging as the world struggles to contain the spread of COVID-19. This may be the result of the social distancing measures mandated by governments worldwide.

An ensemble learning strategy for panel time series forecasting of excess mortality during the COVID-19 pandemic

“…The neural network autoregression (NNAR) refers to single hidden layer networks using the lagged values of the time series as inputs and automatic selection of parameters and lags according to the Akaike information criterion (AIC) [45] . In the NNAR model specification, we considered the last observed values from the same season as the inputs to capture the seasonality patterns and to use a size equal to one, because we have one attribute without a regressor, and by way of improvement, we used 100 networks to fit the different random starting weights and then averaged them out to produce the forecasts.…”

Forecasting the Impact of COVID-19 Epidemic on China Exports using Different Time Series Models

“…For instance, Anastassopoulou, et al, (2020) used the mean-field Susceptible-Infected-Recovered-Dead (SIRD) model to demonstrate the magnitude of spread, contagion, and mortality rates COVID-19 in China during the early period of the pandemic. Other studies conducted by Safi and Sanusi (2021), Petropoulos, et al, (2020), andPerrella, et al, (2020) used the exponential smoothing model to forecast the spread and recovery rates of the coronavirus.…”

An Ensemble Learning Strategy for Panel Time Series Forecasting of Excess Mortality During the COVID-19 Pandemic

“…The Neural network autoregression (NNAR) refers to single hidden layer networks using the lagged values of the time series as inputs and automatic selection of parameters and lags according to the Akaike information criterion [47]. In the NNAR model specification, we considered the last observed values from the same season as inputs to capture the seasonality patterns and use size equal to one, because we have one attribute without regressor, and for improvement, we use one hundred networks to fit with the different random starting weights and then averaged for producing forecasts.…”