Prospa de-risks loan book by mapping borrower relationships

Prospa, an online lender for small business, has used hackathons and graph database technology to map out complex relationships between directors and entities applying for loans.

The aim is to de-risk the loan book by having the necessary context at hand to assess loan applications quickly, manage exposure to entities or people, and predict potential loan distress.

Head of Data and Analytics Jin Foo told the iTnews Podcast that these efforts are underpinned by a data science function assembled over the past three years, comprising full-time staff and PhD candidates from Macquarie University.

“We originally were just connecting data for other people to use, but …more and more the business realised that this data isn’t just good for reporting, it’s an asset we could use to understand our customers, understand what they want, or perhaps what their profile is and what’s the right [loan] product for them,” Foo said.

Foo was given an initial challenge of identifying unknown or hidden exposure to individuals that have – or are involved in – multiple businesses with loans.

Business owners often control multiple entities through trusts and shell company structures that hide relationships between borrowers. A director might appear on one loan application while controlling ten other companies—each with different borrowing histories.

Pre-dating the data science function and capability, Prospa had been manually verifying customer identities and relationships across multiple relational databases and systems.

“We had very complex SQL queries across CRM data, loan origination records, loan servicing information, government business registries and credit bureau data, that would try to account for every possible combination of person and business,” Foo said.

“It gets very complicated, very fast.”

Before joining Prospa, Foo worked in grocery retail, where he used graph database technology through the pandemic to manage stock levels in warehouses, and understand key dependencies that could impact distribution.

Collaborating with Macquarie University, “five or six versions of a customer-based graph” database was set up “within a few weeks” to test the efficacy of the technology.

One of these was built on open-source Neo4j’s free Community Edition.

Prospa ended up using this for a year, internally hosted, before usage levels and operational overhead made it more feasible to run in Neo4j’s managed cloud service instead.

The first experiment around complex relationship mapping, run through a hackathon, produced a result, linking a non-beneficiary shareholder to multiple entities.

“We managed to find an example of a customer who had maybe 30 different loans for 15 different companies, all with us, but we hadn’t noticed the linkages,” Foo said.

The relationship-mapping capability came at a time when interest rates were rising after years of historic lows, so it was additionally valuable to identify individuals that could face distress due to the economic changes.

“We used the insights from graph database technology to alert customers to potential distress,” Foo said.

AI extensions

Prospa is testing the potential for business users to access customer relationship data in Neo4j using Microsoft Copilot for natural language queries.

The intent is for business users to be able to use prompts, such as “Show me customers at risk of arrears in the next 30 days who we haven’t contacted in 60 days.”

Once again, Foo proved this use case through a hackathon. “You can see my strategy here: everything I want to try, I use a hackathon to get the buy-in that I need,” he said.

“I need to show business value and build a business case, and if I can build enough excitement at a hackathon, that helps a lot.”

Prospa is also already experimenting with agentic artificial intelligence (Agentic AI) as a way to automatically assess and flag issues in large amounts of documentation that often accompany a loan application.