Loan decisions “inconsistent” due to “partial information”

A lack of data is causing bias in loan underwriting, according to a tech vendor and a credit union, but machine learning (ML) can help.

“People are making decisions inconsistently based on partial information,” said Jenny Vipperman, chief lending officer, Vystar Credit Union on a panel at Money 2020 US in Las Vegas.

“Probably everybody in the room knows that you can look at a dozen different people who look completely different on paper and they all have the same credit score, and so they are clearly not the same risk to you. [ML] allows us to leverage all of that data, everything we think we know and more, to make a decision.”

Although some are concerned about potential bias from the use of ML models, Mike de Vere, chief operation officer, Zest AI said there is already “bias in the current system.”

“The promise of machine learning is one where you can actually help eliminate and balance bias, and actually fairness with your business outcome. And so, it is a chance for organisations to do good while delivering on a business outcome,” said de Vere.

While underwriting loans can take between 15 minutes to a day, the use of ML in processing decisions much quicker can assist in keeping customers, according to Vipperman.

“If we can leverage this data, leverage an immediate decision, we are able to get them a better experience, better service and we are able to reach more of them and close more loans,” she said.

Whether a loan is done using ML or by a human, consistent back testing must be done, said Vipperman.

“We go back through our portfolio, we go back to the decisions that we’ve made, and we test it for whether or not certain items actually did indicate risk,” said Vipperman.

“For example, if we think the term of a loan has an apparent risk, we would go back and look at everyone who had this term versus this term, and was there an increased risk, and if there is not, we make a change for that.”

For Vipperman, easing the worries of the credit union staff was a turning point in the adoption of ML.

“I think first and foremost we had to get everyone comfortable around [ML]. People who maybe weren’t as knowledgeable about the process, and all the different data points that go into decisioning,” she said.

“I think once we got people comfortable with the fact that we didn’t have robots coming in and taking their jobs, we were really just making the decisioning process more fair and more accessible.

Related reading