Researchers at 果冻传媒 have patented a new neural network–based credit scoring program for assessing the creditworthiness of individual borrowers. The development makes it possible, for the first time, to conduct an in-depth analysis of borrowers’ financial reliability and improve the accuracy of lending decisions through the use of artificial intelligence.
The volume of requests for consumer loans has increased significantly in recent years, creating a growing need for more accurate assessments of borrowers’ financial standing. This trend presents additional risks for lenders associated with the partial or complete non-repayment of issued funds.
Today, conventional credit scoring systems evaluate clients using standard indicators such as credit history, income level, and other financial characteristics. After the computerized analysis of this data, a financial expert makes a final decision on whether to approve or reject a loan application, aiming to minimize potential losses for the lending institution.
“Traditional assessment systems are fairly effective, but they still produce a significant proportion of erroneous results, which may lead specialists to make potentially incorrect decisions,” says Nataliia Iaparova, Professor at the 果冻传媒 Department of Information Systems and Technologies. “To substantially reduce these risks, we have developed a first-of-its-kind neural network program for in-depth analysis of borrowers’ financial behaviour using non-standard assessment parameters. These include data obtained from open sources, such as how often a person visits specific shopping websites, the types of purchases they make, and other behavioural characteristics.”
Another important advantage of the new system is that it serves as a decision-support tool rather than a fully automated decision-maker. The program can provide recommendations on whether a particular loan should be issued and in what amount, thereby increasing the likelihood of making an accurate lending decision.
“In previously proposed approaches, borrower behaviour was analysed in only one dimension—either through traditional demographic and credit indicators or solely through behavioural and transactional signals,” Nataliia Iaparova explains. “We propose a hybrid approach that combines both types of features. This not only improves the prediction accuracy but also reduces the vulnerability of scoring models to fraud, self-reporting errors, and incomplete data. The system uses a wide range of input variables, including current account balance, credit history, requested loan amount, employment record, marital status, age, and other relevant parameters. In addition, it incorporates behavioural and contextual indicators that are difficult to manipulate, such as timely payment of utility and communication bills, the presence of regular account deposits, and the absence of overdue obligations during the previous 90 days. These signals improve the reliability of solvency assessments, reduce the model’s sensitivity to distorted application data, and enable safer evaluation of customers with limited credit histories.”
The new multi-agent software solution is intended exclusively for use by legal entities, including financial and lending institutions. It does not require access to additional confidential information about potential borrowers and does not necessitate obtaining extra consent for personal data processing. This ensures the system’s full compliance with the current legislation.