The Effectiveness of AI Methods Compared to Statistical Analysis and Data Modeling in the Construction of Scoring Models

In the work, we made a comparative analysis of the effectiveness of two analytical cultures, statistical analysis and data modeling (DMC) and algorithmic AI analysis (AMC) in scoring models for calculating LGD (Losses Given Default) – the bank’s percentage share of credit losses (in case of non-payment or delayed repayment of the loan). Analysis based on simulation data generated on the basis of probability distributions, which were obtained from empirical data in a credit risk study at a Polish bank. In the article, we proposed variable selection methods in the DMC and AMC models by analyzing the importance of variables (drop-out and permutation). The proposed methods of variable importance (used for the variable selection) in both cultures showed discrepancies. Each considered model, regardless of the number of observations, indicated different information about the importance of the variables.