This topic keeps popping up here.
Yes, Upstart is FinTech … but FinTech is not Upstart. Same for Upstart and AI …
Lumping Upstart with all Fintech companies and all AI applications and the related racial biases is not accurately looking at the whole picture.
Upstart AI and ML models are scrubbing current data and actions looking for correlations to the ability to pay back a loan or creditworthiness that we humans can’t see. By looking for NEW predictive data points to include in the model vs basing decisions on past (potentially racially biased data) the models become less racially biased.
Do AI models have the potential to be racially biased? They most certainly do, especially when using historical data points that were biased to begin with. But when this is monitored and addressed, they can open up the arena to those who have traditionally been under-served and discriminated against.
Dave Girouard, Upstart CEO stated at a congressional hearing in May “The concern that the use of AI in credit decisioning could replicate or even amplify human bias is well-founded,” and “It’s very reasonable to be concerned about AI because it’s a very sophisticated system and it does have the potential to introduce bias or unfairness. Our answer to that concern is very rigorous testing.”
Upstart also submits the results of the internal testing for bias every quarter to the Consumer Financial Protection Bureau (CFPB).
The data has proven this to be successful to a certain level. Upstart’s AI lending model (per UPST in 2020):
- approved 30% more Black borrowers than “a traditional model” with interest rates that were on average 11% lower
- approved 27.2% more Hispanic borrowers with an average of 10.5% lower rates
This in no way proves that racial biases don’t exist in the UPST models, in fact UPST believes biases can and do occur, thus the effort to monitor and address. It does show that UPST models are less biased than traditional models and are heading in the right direction.
I agree with Jabbo’s take on the demand letter from the NAACP and SBPC in July 2020 and have read the report released in Feb 2020 which led to the letter. At the time of the report, Upstart had already demonstrated that alternative credit data they use had helped them approve 27% more loans than traditional models since late 2017, and that the APR on those loans is on average 16% lower.
Upstart agreed that while their models were already more inclusive than more traditional models (including those used by other Fintech companies), they would increase the scrutiny of the models even further because one of the elements of their objectives to remove racial biases from lending.
Upstart stopped using SAT scores and never zip codes in their models – historical data points proven to be racially biased and used in FICO scores (which are narrow in scope and inherently backward looking). Upstart doesn’t disclose what correlations the data scrubbing has unearthed nor the data points in their constantly improving models, but other companies in China using AI/ML for personal loans for years have shared that use less obvious predictive data points such as how fast the user enters data on the phone, level of phone battery charge, as well as timing of the repayments relative to the due date – data points that I would find hard to think are racially biased.
I believe Girouard has installed a corporate culture to address concerns head on vs deflecting and defense which is so often prevalent in the lending institutions. And I believe this culture will serve them (and the stock price) well going forward.
- Long UPST (though late to the party)
- Pleased with the Giants performance against the Dodgers this past weekend
- Starting to think Fidelity stock news is becoming as relevant as Yahoo news