Conflict of user accuracy requirements and regression model indicators

The paper investigates the relationship between different groups of quality indicators of the re-gression model. Three groups of indicators are considered: statistical indicators, indicators of approximation accuracy in accordance with the requirements of technical applications, and the-oretical indicators in relation to regression coefficients. The statistical indicators are those that are the most common in applied research (although they are not enough for a reasonable as-sessment of the regression equation quality), namely the residual variance and the multiple cor-relation coefficient. Approximation accuracy estimates are those that are most often put for-ward as adequacy criteria in technical applications: average relative deviation of the model from experimental data, maximum relative deviation, average absolute deviation, and maxi-mum absolute deviation. In addition, the difference between the values of the correlation coeffi-cients and their estimates is considered. In order to cover the possible variants of the conditions under which the model is built, the research was conducted under the following conditions: a) compliance with all the prerequisites of the regression analysis; b) compliance with all the pre-requisites of the regression analysis in the presence of large dispersion (large dispersion of re-producibility); c) the presence of heteroskedasticity; d) the presence of outliers (with different locations and different numbers in the training sample). It was established that when the prereq-uisites and assumptions of the regression analysis are met, the difference when using different methods is insignificant from an applied point of view for the indicators of all groups. With large dispersion and heteroskedasticity, methods with the aim of obtaining the best relative in-dicators give a significant improvement of «their» characteristics with a significant deteriora-tion of all other characteristics. Recommendations for the selection of methods depending on the characteristics of the data and requirements for model characteristics have been developed.