This research describes the problem of classification method selection for credit scoring. Credit risk is one of main risks of financial institutions. That’s why researchers and analysts are constantly looking for better credit scoring model creation. Nowadays, a number of different features, such as information about the clients: his habits, previous credits, bank accounts’ information, insurance and other, can be included in credit scoring models. In recent years, deep learning become one of the most popular machine learning methods due to its performance. One of deep learning methods is convolution neural network (ConvNet), which learns all aspects directly from the data, by creating network layers, which in abstract level reveals the main features of the category. This method is suitable to accommodate large number of features and have good accuracy results, what’s why such companies as: Google, Facebook, Microsoft, IBM is using it for different classification tasks. Three real-world data sets from UCI repository are obtained in the experiments for credit scoring problems. Different ConvNet architectures were created random for each data set, based on the logic of the wrapper method (to optimize the accuracy). These results are compered in three ways: with each other; with conventional machine learning methods; with the works of other authors. The experimental outcomes reveal the further development of the ConvNet is expedient in a larger number of features and data sets, either in the combinations of different voting of classifications algorithms (including the ConvNet method).