Abstract: In the article has been conducted a research aiming increase of classification result stability of commercial bank’s debtor creditworthiness with usage of boosted decision trees and neural network algorithms due to the use of stratified sampling. It is proposed to improve the classical procedure of stratified sampling by taking into account not only the target variable, but also the most significant predictors of the model when forming the control subset.

Experimental calculations to test the proposed hypotheses were carried out using the program packages LGBM and H2O on the data of international consumer finance provider Home Credit. In the article checked and confirmed that the use of stratified sampling in the process of forming a control subset during training of machine learning models makes possible to increase their stability and accuracy of forecasts on new data sets.

As per the achieved results, the authors’ approach of stratified sampling during forming a control dataset by target variable and the most significant characteristics of a model demonstrates a higher average accuracy for boosted decision trees on the test subset compared to the standard stratified sampling algorithm and random selection.