Comparing Econometrics Approach Vs. Deep Learning Approach in Forecasting Covid-19 Infections and Deaths Horizon in Egypt

Recently, researchers have applied the Artificial Intelligence, especially deep learning in forecasting, instead of the traditional econometrics’ methods. They argue that deep learning approach can improve the forecasting results by reducing the forecasting errors and saving time and cost. However, there is no empirical evidence for that, since there is a lack of research comparing the forecasting performance of these approaches. Consequently, the aim of this paper is to examine the latter argument and to identify the best technique to be used in forecasting the Covid-19 pandemic and others. In fact, the Covid-19 pandemic is defining a global health crisis, which is the hugest the world has faced since World War II. In addition to being threatened by GDP decline and income losses; fears of fetal effects of this epidemic makes it critical to predict the potential spread and identify the best techniques to be applied for that purpose. To achieve the aim of this study, two different methods of forecasting, namely an econometric approach named Autoregressive-Distributed Lag (ARDL) and a deep learning model named Long Short-Term Memory (LSTM) are utilized to forecast the number of daily cases and deaths of Covid-19 in Egypt (March 2020 - March 2021). Consequently, the contribution of this paper is twofold; first, investigating the best way of forecasting the Covid-19 epidemic especially, and therefore real-life phenomena in general, second, assessing the impact of mobility on the incidence of the pandemic in Egypt. The results revealed that the LSTM method shows a slightly better forecasting performance even without using mobility data. statistical AI the ARIMA and the Nonlinear Autoregressive Artificial Networks (NARANN). the reported cases 1 May one To assess and compare the performance of the two models, the the following measures; Mean Absolute Error (MAE), Root Mean Square Error (RMSE), R2, Deviation Ratio (RD) and Coefficient of Residual Mass (CRM). The study concluded that the NARANN outperformed ARIMA as the latter resulted in an increasing error rate over time from 3.08% for the one-day ahead forecasting to 29.48% ten-days ahead. Conversely, the NARANN results showed an absolute percentage error fluctuating between 1.12% and 4.89%.