MODELING THE AMOUNT OF INCOME FROM THE SALE OF THE UTILITY COMPANY'S PRODUCTS

Abstract

The task of increasing the reliability of estimating the level of net income from the sale of products of a utility company is studied. The dependence of the amount of net income from sales of products on such factors as cost of sales and other operating income is analyzed. The presence of multicollinearity between the factors necessitated the construction of regression models using the least squares method for each factor. To build a stepwise model, the paper uses a transformation based on the decimal logarithm. We also built models using the least squares method after applying the principal components method to the factors. To test the internal parameters of the regression models built using the least squares method for significance, we used the value of their p-values. All models where the primary data were transformed using the principal components method failed the significance test. After checking the internal parameters in order to find the best model, the obtained regression models were compared using the following indicators: coefficient of determination, Fisher's criterion, and average approximation error. As a result of the comparison, the logarithmic model was recognized as the best single-factor regression model by the least squares method. To solve the problem of increasing the reliability of estimating the level of net income from the sale of products of a utility company, the paper also considers models based on machine learning. Regression models based on k-nearest neighbors and a binary decision tree were built. To build models using machine learning methods, the initial data were standardized and transformed using the principal component method. When comparing the resulting models, it turned out that the two models based on machine learning methods have a smaller approximation error than the models based on the least squares method. The best regression model for estimating the level of net income from the sale of a utility company's products was the binary decision tree model. The coefficient of determination of this model was 0.987, and the value of the average approximation error was 0.91 %.