Sberbank and Skoltech: Method of training neural networks for the banking sector

The main articles are:

• Neural networks (neural networks)
• Artificial intelligence in banks

2025: Introduction of Neural Network Training Method for Banking

Scientists from Sberbank and Skoltech have increased the accuracy of neural networks for the banking sector by 20%. This was announced by the Russian Scientific Foundation (RNF) on February 28, 2025.

𝓲

Фото: RSF

Scientists have developed a method for training neural networks, thanks to which algorithms can take into account both local and global levels of banking data. This will help optimize business processes, improve the security and quality of customer service in banks.

Banks use artificial intelligence (AI) algorithms to process and analyze data and predict a range of processes. For example, AI allows you to analyze financial transactions - money transfers between organizations and individuals - for which you can predict the bankruptcy of companies and the risks of non-repayment of loans.

In addition, neural networks help to automatically identify cases of fraud in transactions that are not typical for the user, for example, large debits from accounts, as well as in calls from suspicious numbers. AI also helps banks select individual offers for customers depending on their needs. At the same time, to solve different problems, you need to analyze very different data sets - from the local level (a sequence of individual banking operations in a short period) to the global one (the entire history of customer transactions). However, existing algorithms focus only on one of these levels, which is why not all problems are solved equally efficiently.

Researchers from the Artificial Intelligence Laboratory of Sberbank and Skoltech (Moscow) have proposed a method for training neural networks that are used in the banking sector for tasks for processing event sequences. The authors divided such tasks into three types: global, local and dynamic. Global ones require an assessment of some general characteristic of the sequence, which almost does not change over the period of time in question: customer age, solvency, satisfaction with the bank's services. Local ones rely on a characteristic that constantly changes in time, for example, the definition of fraudulent transactions. Local tasks require that the neural network can quickly respond to sharp changes in client behavior, for example, to identify a change in country of residence.

On all of the above tasks, the researchers tested a large set of models used to analyze sequential data. Based on the results, the authors developed a completely different analysis method. It consists in the fact that the analysis takes into account external context information, that is, data on other clients, especially those who are similar in a number of features to the analyzed. This helps to take into account various global trends. This approach improves the quality of models on all proposed tasks, in some cases by a margin of 20%.

Most of the tasks that we worked with before the start of this study could be attributed to global, but we tried to work ahead of the curve and find algorithms that will cope well with local productions. Surprisingly, as of February 2025, most of the tasks that arise before us are rather local. It turned out that the practical need had just appeared, and we already have a good solution ready. In my opinion, this is one of the main advantages of the work, distinguishing it from most journal articles on artificial intelligence, which at the time of publication are already a little outdated, "said Andrei Savchenko, Doctor of Technical Sciences, Scientific Director of the Laboratory of Artificial Intelligence of Sberbank.

One of the special properties of neural networks is versatility, the ability to adapt to different problems without additional costs. In our work, we were able to describe a large set of tasks and offer solutions that do a good job with everyone, including in the event of a change in user behavior over time. Separately, I am proud that it turned out in the model to take into account the behavior of similar users, which led to a further increase in the quality of the model. The work does not end at the publication, and then we plan to use the method for new types of data, to increase the resistance of neural networks to anomalies, - summed up the head of the project supported by a grant from the Russian National Research Fund, Aleksei Zaytsev, candidate of physical and mathematical sciences, head of the Skoltech-Sberbank applied research laboratory at the Skoltech AI Center.