Nonparametric Tests for Comparing COVID-19 Machine Learning Forecasting Models

Klyushin, D.

The rapid spread of the COVID-19 pandemic has revealed an acute problem associated with forecasting the development of epidemics in different countries and assessing the accuracy of these forecasts. The most widely used models for predicting viral outbreaks are the SIR and SEIR models and their modifications. The complexity of predicting the development of epidemics using these models is associated with the large uncertainty of the parameters included in these models;therefore, many researchers have attempted to apply machine learning methods for forecasting. The objects of our study were the Grey Wolf Optimizer (GOW) methods, as well as two machine learning models (MLP multilayer perceptron and ANFIS adaptive network fuzzy inference system). Ranking the models in increasing order of their accuracy is not difficult, but in many cases it makes no sense, since the model predictions in the statistical sense differ little from each other (homogeneous) and therefore such models can be considered equivalent. The chapter describes an effective nonparametric tests both traditional and new for comparing the results of time series prediction, which makes it possible to assess their homogeneity and identify groups of forecasting methods whose forecast results differ statistically significantly from each other. The results of comparison of forecasting machine learning models for different countries are presented. The effectiveness of the applied method of nonparametric testing of the homogeneity of models is shown. © 2021, The Author(s), under exclusive license to Springer Nature Switzerland AG.