Effectiveness of Variable Selection Methods for Machine Learning and Classical Statistical Models

In line with new international financial supervision directives (IFRS9), banks should look at a new set of analytical tools, such as machine learning. The introduction of these methods into banking practice requires reformulation of business goals, both in terms of the accuracy of predictions and the definition of risk factors. The article compares methods for selecting variables and assigning "importance" in statistical and algorithmic models. The calculations were carried out using the example of financial data classification for loan default. The effectiveness of various machine learning algorithms on selected sets of variables was compared. The results of the analyzes indicate the need to revise the concept of the "importance" of a variable so that it does not depend on the structure of the model.