Everyone knows Dun & Bradstreet from its credit reports. But what is really behind those credit reports? For 30 years, we have been busy developing our scorecards. And we don't take that lightly. Our goal? To predict which companies might go bankrupt or pay late in the future, all based on the abundance of data and information we have collected in D&B's huge database.
During this time, we invested heavily to improve the predictive power, relevance and logic of our scoring systems. These efforts have paid off, as now our scoring systems perform automated risk assessments in as many as 41 countries around the world. All types of companies are assessed by these systems, from small independents to large international corporations. This is all done regardless of whether financial data is available.
Within D&B, scorecards have been developed with the latest machine learning techniques for years. Recently, for Belgium, we added Artificial Intelligence.
See: The secret behind our annually improving scores
Why are credit ratings so important?
Credit ratings as we know them today were developed during the last century in the US, with the primary aim of facilitating consumer credit decisions. Whereas previously decisions were made solely on the basis of a manual credit rating, scorecards support this process with numerical ratings that reflect the creditworthiness of a prospect or customer. This allows better identification of risks, allowing lenders to make faster and more accurate decisions on lending and credit. Based on the scores, some companies went even further and already automated whole chunks of the assessment process.
The rise of AI in credit ratings
Back in the last century, our industry began exploring how artificial intelligence ('AI'), in the form of Neural Networks (or Artificial Neural Networks, or ANNs for short) could be used to develop credit scores. Those ANNs are computer models inspired by the structure and functioning of human brain cells. They are made to learn and recognise complex patterns in data. The starting idea was that the "power" of these systems could increase the predictive power of credit ratings. Unfortunately, it was then found that the additional predictive power of these ANNs was not great enough. Especially considering the complexity and explainability of these types of models. As a result, the use of ANNs for credit assessment disappeared more into the background.
However, with the rise of social media and companies in the 'new economy', interest in applying AI for credit assessment grew again. A not insignificant part of these companies' investment in AI, is making it explainable ("xAI", or Explainable Artifical Intelligence).
Interesting read: xAI for tomorrow's credit scoring
Against the backdrop of this evolution, we began to explore (again) how we could use AI to develop scoring systems. In doing so, we had two main goals in mind: first, we wanted to increase the predictive power of our scoring systems, and second, we wanted to ensure that our system would be manageable, transparent and understandable.
AI in our credit ratings - more revenue, less risk
In 2019, we achieved both goals with the development of our first-ever credit scoring system integrating an AI layer. In doing so, we did not want ANNs to directly generate our scores. Instead, the AI was embedded between two layers: a 'logic layer', which provides direction and guidance for the AI, and the traditional scorecard layer, which is based on the outcome of AI.
This approach, which we call 'AI Layered Scoring', allows us to further sharpen credit risk assessment. We have become even better at identifying companies that cause a credit loss against those that are bona fide.
What does this mean in practice?
A concrete example, based on a system we developed for a client, makes this clear. The objective of this system was to predict defaults as accurately as possible. Suppose we wanted to see 60% of defaulters over a given period, arriving well in advance. With traditional machine learning for score development, about 24% of customers got a "low" score, to achieve this 60%. With our new method, we only need to give 5% of customers a low score to identify this 60%.
The difference between the two (24% vs. 5%) shows that we were able to achieve a significant improvement here: 19% more companies can do turnover without increasing the risk of default. More turnover, less risk, and thus better returns.
These final scores help you in deciding who you do business with. With an abundance of data, it is extremely important that ease of use is key. We understand better than anyone that the willingness to take risks varies from one organisation to another. Our custom scoring module offers the ideal solution for this. Curious to learn more about AI Layered Scoring and how it can help your business? Contact us below.