The challenge
In many companies, including Microsoft, credit risk management was done by an attribute scoring system. It worked in the following way, “we collected attributes from the customer (eg, financial statements, country, industry, number of employees), from our historical relationship with that customer (eg, payment performance, years of relationship), and from third parties, most commonly credit bureaus like S&P and Dun & Bradstreet (eg, credit rating and paydex score),” explains Dennis Crispin, Group Manager, GFS.
“We then assigned a scoring rubric to each one of these attributes, and then added up all the scores. Customers with high scores would be granted credit, while customers with low scores were not. As an example of a rubric, consider a country: a customer in Libya would get a low score on the country attribute, while a customer in Switzerland would get a high score.”
These rubrics were created by a combination of experience and expertise, and work well enough when the overall market conditions are somewhat stable. Maintaining these rubrics was a large undertaking when the market was stable, and even larger in the face of severe economic challenges such as the one caused by COVID-19.
Microsoft noticed that, on average, more than 99% of the credit requests were being granted with no modifications, and that, because of the volume, it was taking the company more than one business day to approve credit. Overall losses were low both in absolute terms and compared to the industry.
“Putting this all together, we realised we had a system that relied a lot on manual work, approved a large proportion of the cases despite this manual work, effectively delaying sales, and that was costly to maintain and operate,” says Lucas Mayer, Group Manager.
The solution
Microsoft looked for a solution that would speed up the credit decisions, while reducing manual labour and at least maintain the current performance.
The company realised that the solution would need to address the following points:
- Small number of inputs.
- Strong preference for inputs that are machine-readable and require no human touch.
- Credit losses on par with the performance achieved by human credit managers.
- Fast decisions for simple situations, hopefully in seconds.
- Self-evolving, reacting to changes in market conditions without human interference.
“We created a cross-functional team with people from technology, analytics, process, audit and credit risk to suggest solutions that would fulfil the requirements above,” says Crispin.
After trying out several different ideas, Microsoft decided to implement an AI-based solution. The company’s analytics team had several successes in the collections area and a lot of experience with the data required for credit and collection decision automation, so set out to create a model that would attempt to replicate or improve on human performance.
The solution uses a machine learning model created by Microsoft Research called LightGBM. Using its Azure platform for cloud-based computing, it ingested data from its data warehouse and from Dun & Bradstreet using Azure Databricks and processed large amounts of data using Azure Machine Learning. With this set-up, the company successfully trained a machine learning model that achieved human-level performance in just a few weeks.
Best practice and innovation
Microsoft partnered with Dun & Bradstreet to automatically ingest a significant amount of structured customer information automatically, and prioritised structured information over unstructured information. The intended goal was to see if a machine learning model using only structured data was able to achieve the same performance as a human credit analyst that used structured data, unstructured data and judgment.
Key benefits
- Much faster credit approval.
“Although a one-day improvement may seem small, the amount of revenue supported by this process is huge, over US$40bn. Using a 2% cost of capital rate, this adds up to US$2.2m per year in cost of capital savings,” says Edward Lu, Group Manager.
Listen to podcast
