Combined Feature Selection Scheme for Banking Modeling

Machine learning methods have been successful in various aspects of bank lending. Banks accumulated huge amounts data about borrowers over the years application. On one hand, this made it possible to predict borrower behavior more accurately, on other, gave rise problem a redundancy, which greatly complicates model development. Methods feature selection, allows improve quality models, are apply solve problem. Feature selection can be divided into three main types: filters, wrappers, and embedded methods. Filters simple time-efficient that may help discover one-dimensional relations. Wrappers effective because they account for multi-dimensional relationships, but these resource-consuming fail process large samples with many features. In article, authors propose combined scheme (CFSS), first stages use coarse final — wrappers high-quality selection. This architecture lets us increase reduce time necessary samples, used development industrial models. Experiments conducted by four types modelling tasks (survey scoring, behavioral customer response cross-selling, delayed debt collection) shown proposed method better than classical containing only filters or wrappers.