Mikhail P. Bazilevskiy

Irkutsk State Transport University

Previously, the author proposed modular linear regression models containing as regressors modules of deviations of the values of explanatory variables from unknown coefficients. An algorithm for their exact estimation using the least absolute deviation and an algorithm for approximate estimation using the least squares method is known. Software products implementing these algorithms have not been developed until today. This article is devoted to the description of the software package developed by the author for evaluating modular linear regressions (PC MODULIR-1). In it, when evaluating modular linear regression using the method of smallest modules according to the specified settings, a mixed-integer 0-1 linear programming problem for the LPSolve package is automatically generated. And in the case of approximate estimation using the least squares method, a complete search of all possible model variants is carried out and the best modular regression with all coefficients significant according to the Student's t-test is selected. The problem of modeling the freight turnover of railway transport in the Zabaykalsky krai was solved with the help of PC MODULAR-1. The coefficient of determination of the modular regression constructed using the least squares method with five explanatory variables was 0.94, which is about 4 times higher than that of traditional linear regression. At the same time, all the coefficients of modular regression turned out to be significant according to the Student's t-test. It is shown how the constructed modular regression can be interpreted.

modular regressions, software, least absolute deviation, least squares method, coefficient of determination, Student's t-test, cargo turnover