It’s
been the Catch 22 of credit: Consumers who aren’t continually active
users of credit struggle to get approved for new credit accounts.
For
years, consumers who didn’t have a recent track record of credit cards,
loans or other types of borrowing were either frozen out of getting a
loan altogether, forced to pay higher interest rates or required to turn
to other alternatives, such as pawn shops or payday lenders.
VantageScore
3.0, the most recent commercially available VantageScore model, scores
up to 30-35 million consumers who can’t obtain a credit score under the
older, conventional scoring models used by many financial institutions,
including all mortgage lenders. The recently announced VantageScore 4.0
model, which will be commercially available this fall, applies the
latest in computer-driven, machine-learning technology to drive
additional predictive performance within this population of consumers.
Specifically,
VantageScore 4.0 leverages this technique to generate a more predictive
credit score among those who have not engaged in credit activity in the
past six months, but have relevant information in their credit file
that is more than six months old.
Data scientists at
VantageScore used machine learning to take an even deeper dive into the
credit histories of these consumers who are often overlooked by lenders.
They examined a large number of specific details within credit reports
to identify additional patterns that correspond with good financial
behavior or increased risk.
“These consumers may not be
high-frequency users of credit; and more importantly, have not engaged
in credit activity in the last six months,” explains Nick Rose, one of
the principal scientists for VantageScore. “Their credit use tends to be
sporadic. For example, they might open an account for a few months, and
then not use it again. If you try to score them with the conventional
credit-scoring techniques, you can’t get any meaningful insight based on
their credit usage and patterns.”
Indeed, although it’s
relatively easy to predict the creditworthiness of a borrower who has
had a mortgage for several years, as well as car loans, credit cards or
other installment debt, analyzing and predicting the credit behavior of
consumers with less recent credit histories called for a more technical
solution.
“Machine learning found some 50,000 different
relationships worth considering,” said Rose. “We took those 50,000 and
reduced it to 300 potential relationships we could focus on to determine
which ones were the most valuable and predictive,” Rose says.
For
example, the machine learning approach helped data scientists assign
default rate profiles based on the relationship between the balances on
collection accounts, the ages of the collection accounts and the numbers
of inquiries that those consumers had recently made.
Importantly,
attributes that were discovered with this approach also adhere to
certain standards for inclusion into a generic credit scoring model,
such as ability to translate the attributes into predictive reason
codes.
“The use of machine learning provided a significant
performance boost within a key population,” said Rose. “And equally
important is that we were able to incorporate our findings into the
confines of a regulatory compliant, generic credit scoring model that
lenders can implement into their current systems.”