Context: the company has a lot of corporate clients-borrowers for whom it is necessary to constantly monitor credit risks in order to properly manage the client portfolio. One way to assess possible risks is to mention specific events of the borrower in the news stream (for example, the change of the CEO). News flow analysis can be automated to increase efficiency.
Decision: a news tagging model was proposed based on 17 main possible risks. Using the model, an index was built for each corporate client, demonstrating the real level of risk for all news.
Results:
Integration of the Bank's loan portfolio analysis Department into the business process:
- Reducing the load on the analyst by up to 70% for various risk categories; Improving the accuracy of the Department's forecasts by 15%;
- Increase in the number of detailed reports on the portfolio by 2 times;
- Creating a semantic core for 80% of client companies.
To build the model, we used:
- News stream for the selected borrower company;
- Set of risks set by the Bank;
- Markup of news according to the risks from analysts.
Simulation result:
Model for assessing the probability of risks for a given borrower company based on previous news.
Customer: Finance, Banking
Technology stack: Python, nltk, gensim, PyTorch.