Selection of regional determinants of economic security

Abstract

Subject. The article addresses the selection of informative and significant regional determinants of economic security. Objectives. The aim is to explore statistical and econometric tools to select regional determinants, to build economic security classifiers, using the machine learning algorithms. Methods. The determinants are selected, using the filter-based methods (on the basis of Chi-square, F-criterion, information content criterion), sequential and recursive selection in the opposite direction, exhaustive sampling and selection based on the "random forest" classifier. The above algorithms are implemented in the Scikit-learn and MLxtend libraries. The analytical environment Jupyter Notebook is used as a system providing the execution of feature selection algorithms. Results. I formed a training sample of 90% of the total set of regional economic determinants for 82 Russian regions. The determinants were ranked by significance. Logistic loss and model accuracy were set as a model evaluation criterion. The accuracy of created classifiers was between 78% and 93%. The paper establishes the significance of categorical variables that determine whether a particular region belongs to the resource, investment, or regional profitability cluster. The constructed heat map enabled to establish the degree of linear relationship between the level of economic security of the region and regional determinants. Conclusions. The most significant group of panel indicators to determine the class of economic security is the resource security of regions. The level of resource endowment depends on the volume of consumption and the share of manufactured goods in the GRP. To build a model to classify regions by the economic security level, five features are sufficient; this circumstance was established by automatic set of determinants, using the recursive feature elimination algorithm with cross-validation (RFECV). The implemented wrapping methods of feature selection demonstrated that significant indicators in the model are budget expenditures and the amount of debt per capita, investment risk and net financial result. The best quality of the classifier is provided in the course of an exhaustive set of features, using the K-Nearest Neighbors algorithm.