Now reading:

Adopting AI in credit risk monitoring

Adopting AI in credit risk monitoring

  • 20 November 2019
  • By Nick Luo
  • 5 mins read

Adopting AI in credit risk monitoring

Nick Luo (pictured standing, second from left) is a Data Scientist with the OCBC AI Lab under Group Customer Analytics & Decisioning, and the key person behind the Bank’s auto news-scanning AI model developed for the Wealth Management team. Hear what Nick has to say about the project and how it has improved efficiency.

Q: Could you tell us more about the project and its objective?

A: To put it simply, we were tasked to build an AI model that will monitor and interpret news related to bond investment products, to assist our Wealth Management team in curating relevant market updates to be conveyed to our bond customers.

The Wealth Management team recommends bond investment products issued by over 400 bond issuers to retail customers, and fluctuations in credit risk of bond issuers have to be conveyed to customers accordingly as per the Monetary Authority of Singapore (MAS) fair dealing act. There was a pertinent need to increase efficiency in monitoring such market highlights due to fast and frequent changing market, so as to curate and communicate news to customers in a timely manner.

Q: What was your role in this project?

A: I was responsible for designing and developing the AI model, that was tasked to read, analyse and classify each piece of news as positive, neutral or negative, from a credit risk perspective.

Q: How were the rest of the business involved?

A: Multiple stakeholders were brought onboard, with approximately 25 project members joining the project team and Group Operational Excellence (GOE) taking lead as the main coordinator and project manager of this initiative. Besides the AI Lab, which provided the technical expertise in building the AI model and dashboard, Treasury stepped in as knowledge experts on classification of news, while Wealth Management provided insights on dashboard design and selection of internal and external customer communication channels. Other stakeholders such as Legal and Compliance, Brand & Communications and Business Transformation were also involved at different stages of the initiative.

Q: That sounds intensive. How long did it take for this AI product to materialise?

A: It took approximately 6 months, from conception, vendor selection, proof-of-concept prototyping, to launch in April 2019. We adopted an Agile product methodology, opting to build a working AI model fast, and then subsequently enhancing the model through continuous feedback. This enabled us to shorten the project timespan by a significant amount.

Q: What sort of challenges did the project team encounter?

A: Finding sufficient, good quality data, was a challenge at the early phase of the project, as public news sources were of poor quality and contained too much noise. Eventually, we had to work with a vendor who provided us with over 4 years’ worth of historical news data to train the AI model.

When assigning a sentiment score to articles, it could also be rather subjective, as different parties may hold different opinions for the same piece of news. By training our AI model using news articles tagged by our Group Treasury research colleagues, we were able to incorporate our house views into the model for analysis with higher accuracy moving forward.

Q: That sounds really exciting for the business!

A: Yes, and in fact, we have plans to extend the scope to unit trust, equities or structured products news analysis in the future. This was a successful example of utilising unstructured data, such as images, speeches and news articles, in project application - something which has only been made possible through AI. As one of the first of such projects rolled out internally that has shown great potential, I’ve no doubt that more of such projects will follow in its success.