Forecasting the number of cases in pandemic with machine learning

The coronavirus initially appeared in the city of Wuhan in the Republic of China in December 2019. It was declared as a pandemic by the World Health Organization on March 11, 2020. Since the pandemic started to threaten countries' public health, measures have been taken by the countries to control the number of cases and deaths. Measures such as quarantine, curfew and periodic restrictions, closure of eating and drinking places and social areas for a while form the basis of precautions. The COVID-19 pandemic, which has been affecting the whole world for more than a year, is still effective and remains serious. Although vaccine developer companies have developed various vaccines, the active cases and death rates have not fully extinguished the severity of the pandemic due to the lack of sufficient vaccine production and the optimal distribution of the vaccines produced. During the pandemic period, planning the future days by the decision makers is of great importance in terms of public health. Forecasting the number of active cases in the coming days is an extremely helpful foresight in order to staff and material planning for hospitals, logistics and supply planning, vaccination scenarios, construct vaccine distribution models. In the literature, many COVID-19 active case predictions have been conducting using different methods such as compartmental models, derived compartmental models, machine learning algorithms, and time series analysis. In this study, the United States and Turkey's number of active cases in the coming days were forecasted by machine learning algorithms. Python programming language was used to implement machine learning algorithms such as Prophet, ARIMA, Linear Regression, Polynomial Regression and Support Vector Regression models. The performances of the algorithms were evaluated using the mean absolute percent error (MAPE), root mean square error (RMSE) and mean absolute error (MAE). As a result, for the methods we tried and the data sets we used, the Prophet algorithm showed the best performance for the United States and the ARIMA model for Turkey.