Filter methods of variable selection for enterprise credit risk prediction

• Špicas, Renatas
• Kanapickienė, Rasa
• Ivaškevičiūtė, Monika

LTReikšminiai žodžiai: Variable selection; Credit risk; Filters; Wrappers; Credit scoring.

ENCredit risk continues to be the major source of risk and possible losses in the activity of credit institutions of any type. In context of fast developing informational infrastructure and modern technology, credit risk modelling practitioners face new challenges. Nowadays loan analysis systems allow analysing wide range of possible independent variables, starting from hundreds to tens of thousands possible credit risk indicators. This article focuses on variable selection techniques for credit scoring models with different types of classifiers, including statistical and artificial intelligent classifier. Authors compare different approaches of variable selection techniques, including filters, wrappers and embedded methods, identifying their advantages, disadvantages and limitation to practical use. As a result of the article, filter methods of variable selection are presented and tested on real world dataset. The article uses the following methods – modelling, statistical analysis and evaluation of corporate data. The aim of the article is provide filter methods of variable selection techniques, which can be used as preprocessing step, before using wrapper methods or selecting variables for credit risk assessing models. Provided methods reduce number of possible variables by removing redundant and irrelevant variables by applying bivariate analysis techniques.The research results will contribute to the development of multidimensional state-of-the-art credit risk assessment and pricing model, appropriate to use in terms of the challenging modern financial market. Different types of feature extraction and variable selection techniques are appropriate for solving different tasks and in deferent situations. Even though filter methods are computationally convenient, but they are not appropriate to optimal set selection. Wrapper methods are most accurate, but face the problem of curse of dimensionality effect. Authors provide filter techniques for variable selection, the reliability of which is tested by empirical research. [From the publication]