COVID-19 has become the largest pandemic in recent history to sweep the world. This study is devoted to developing and investigating three models of the COVID-19 epidemic process based on statistical machine learning and the evaluation of the results of their forecasting. The models developed are based on Random Forest, K-Nearest Neighbors, and Gradient Boosting methods. The models were studied for the adequacy and accuracy of predictive incidence for 3, 7, 10, 14, 21, and 30 days. The study used data on new cases of COVID-19 in Germany, Japan, South Korea, and Ukraine. These countries are selected because they have different dynamics of the COVID-19 epidemic process, and their governments have applied various control measures to contain the pandemic. The simulation results showed sufficient accuracy for practical use in the K-Nearest Neighbors and Gradient Boosting models. Public health agencies can use the models and their predictions to address various pandemic containment challenges. Such challenges are investigated depending on the duration of the constructed forecast.

The global COVID-19 pandemic began in December 2019 and spread rapidly around the world. Worldwide, more than 230 million people fell ill, 4.75 million cases were fatal. In addition to the threat to health, the pandemic resulted in social problems, an economic crisis and the transition of an ordinary life to a "new reality". Mathematical modeling is an effective tool for controlling the epidemic process of COVID-19 in specified territories. Modeling makes it possible to predict the future dynamics of the epidemic process and to identify the factors that affect the increase in incidence in the greatest way. The simulation results enable public health professionals to take effective evidence-based responses to contain the epidemic. The study aims to develop machine learning and compartment models of COVID-19 epidemic process and to investigate experimental results of simulation. The object of research is COVID-19 epidemic process and its dynamics in territory of Ukraine. The research subjects are methods and models of epidemic process simulation, which include machine learning methods and compartment models. To achieve this aim of the research, we have used machine learning forecasting methods and have built COVID-19 epidemic process linear regression model and COVID-19 epidemic process compartment model. Because of experiments with the developed models, the predictive dynamics of the epidemic process of COVID-19 for 30 days were obtained for confirmed cases, recovered and death. For ‘Confirmed’, ‘Recovered’ and ‘Death’ cases mean errors have almost 1.15, 0.037 and 1.39 percent deviant, respectively, with a linear regression model. For ‘Confirmed’, ‘Recovered’ and ‘Death’ cases mean errors have almost 3.29, 1.08, and 0.71 percent deviant, respectively, for the SIR model. Conclusions. At this stage in the development of the epidemic process of COVID-19, it is more expedient to use a linear model to predict the incidence rate, which has shown higher accuracy and efficiency, the reason for that lies on the fact that the used linear regression model for this research was implemented on merely 30 days (from fifteen days before 2nd of March) and not the whole dataset of COVID-19. Also, it is expected that if we try to forecast in longer time ranges, the linear regression model will lose precision. Alternatively, since SIR model is more comprised in including more factors, the model is expected to perform better in fore-casting longer time ranges.