Abstract: The article highlights a study on the search for the most effective approach to pre-processing the characteristics of borrowers in order to improve the accuracy of predicting defaults on credit obligations. Three main ways of providing data to the inputs of credit scoring models are analyzed: the use of the initial explanatory variables without transformation, the conversion of categorical characteristics into a set of dummy variables, binning the indicators with the calculation of the weights of evidence (WOE) for each category.

To obtain conclusions about the systematic impact of these approaches, 10 repeated iterations were carried out with the construction of perceptron-type neural network models based on each of these three methods of preparing input factors. All scoring models were evaluated by a wide range of indicators of integrated and point efficiency.

The results of the experiments showed by almost all criteria the advantage of the methodological approach proposed by the author for preliminary data processing by dividing quantitative variables into categories, ensuring the trend in values of their weights of evidence and observing restrictions on the volume of observations in each group.