Selection of optimal complexity models by methods of nonparametric statistics (on the example of production function models of the regions of the Russian Federation)

The article describes an approach to comparing alternative variants of linear regression models on time series and determining the appropriateness of complicating them (by adding new variables) using several variants of Monte-Carlo methods. The proposed research methods using pseudosampling generation allow taking into account both the effects associated with possible differences of distributions in empirical data from the Gauss distribution and the effects associated with possible nonstationarity of the time series under study. For this purpose, pseudosampling generation is used — time series, which are Gaussian white noise, random walk generation, as well as the permutation test and the bootstrap method. Reliability of the obtained results is estimated using resampling. Applicability of the considered methods is demonstrated by the example of models of investment production functions of regions of the Russian Federation, calculated on the basis of data from the Federal State Statistics Service.