Machine Learning’s Impact on Business Credit Assessment
One of the most wonderful things I’ve learned about using machine learning (ML) is how it requires us to grow and change as a company. As we gain new insights and enhance the Dun & Bradstreet Data Cloud and products, we also must adjust our risk analytics definitions and best practices to accommodate for new and exciting developments.
One excellent example provided by Dr. Mark Seiss, Director, Advanced Analytic Services at Dun & Bradstreet, is how we define our reason codes. For credit risk assessments, companies like Dun & Bradstreet must provide a reason for the assessment in a company’s file, informed by its scores and ratings. These reasons are often codified to help provide standardization and give attribution to the variables that go into creating a score or rating and are aptly named “reason codes.”
These codes are extremely important because they provide transparency to our customers, so that they can trust the assessment and take action accordingly, whether they are using a Dun & Bradstreet business credit file to help assess a company’s risk or looking to potentially impact their own file before they bid on a new contract. I spoke to Dr. Seiss about how machine learning has changed how we approach assigning reason codes at Dun & Bradstreet, and what follows is our conversation.
Tanya: Can you provide an example of an application of reason codes that was influenced by our ML models?
Dr. Seiss: Dun & Bradstreet has a variety of scores and ratings that provide our customers with different views of risks associated with businesses. The D&B Financial Stress Score (FSS) is a good place to start when explaining how our progress using machine learning has changed our approach to reason codes. The FSS assesses the likelihood that a business will file for bankruptcy in the next 18 months. An example of a reason for a high-risk FSS assessment would be the presence of a suit, lien, or judgment on a company, and that reason is one of many that impact the assessment. Previous versions of the FSS were built using scorecard methodology, a highly transparent and intuitive version of logistic regression modeling. Due to the transparent nature of scorecard models, deriving reason codes for the FSS has generally been a straightforward exercise for us — until we started using machine learning.
Tanya: What were the challenges that arose?
Dr. Seiss: ML models often outperform traditional modeling methods, such as scorecards, and can provide significant improvements in predictive insights. However, machine learning models have extra layers of complexity that lend to this improved accuracy. They utilize non-linear relationships and interactions between predictive attributes, and these complex relationships make these models difficult to explain simply. The challenge is to leverage the increased predictive capabilities produced by machine learning models and maintain simple, clear transparency when explaining the reasons for assessments to our customers.
Tanya: What solutions have the Dun & Bradstreet analytics team found that have helped increase transparency?
Dr. Seiss: Research advances in reason codes, also known as variable attribution, have enabled us to improve transparency into modern machine learning methodologies. At Dun & Bradstreet, the Advanced Analytical Service team has been examining recently developed reason code derivation methodologies to determine which method will best provide our customers with the most transparent reasons for, and actionable insights from, the assessment provided by our ML-developed business risk scores. The reason code methodologies we choose will also incorporate past lessons learned from previous reason codes derived for our traditional scorecard models.
Our team is always excited to tackle the challenges presented by new methodologies and models that provide more value to our customers. Machine learning is, and has been, helping us mine the informational gold from the Dun & Bradstreet Data Cloud, the world’s most comprehensive source of business data and activity, more quickly and thoroughly than ever. We look forward to the new analytical challenges it undoubtedly will bring. We expect that these challenges will help us get better and better at turning data into the insights that can help our customers improve performance, reduce cost, manage risk, and transform their businesses.
Tanya: Thank you, Dr. Seiss. I look forward to learning more about how your team continues to use machine learning to provide valuable insights to our customers.
To learn more about how Dun & Bradstreet and our work with machine learning can help your business, please download our eBook Machine Learning for Superior Analytics.