Abstract: The problem of neural networks overfitting is studied in the article.

There is revealed the theoretical foundations of this effect occurrence and highlighted the negative consequences of its arising. It’s carried out an experimental research of the overfitting effect by the example of the problem of modelling of individuals credit risk. To this purpose was built a number of logistic regressions and neural networks of perceptron type of different configurations, which adequacy was analysed on the training and test samples. The results of comparing the accuracy of borrowers classification on the basis of Gini criterion for both samples made it possible to get some new generalizations about the nature of the overfitting effect and recommendations for its prevention. The practical value of the study is the obtained new knowledge on how to avoid of the overfitting effect in scoring models that will ensure stability of credit risk modelling results in the banking institutions activity and will contribute to the sustainability of the financial system at whole.