abhijitd1,
One point I did not emphasize enough in my post:
I believe one major reason why the APR offered to my “superprime” status is so high from the lower-prime lenders/underwriters (LendingPoint, Avant etc) and the APR offered from high-prime lenders (BestEgg, SoFi) is lower, is that higher losses incurred by low prime lenders must be made up for by charging everyone else higher APR to cover the costs of defaults and volatility.
UPST cofounder Paul Gu explains this concept well from a 2015 talk (when he was like 23 years old!):
https://m.youtube.com/watch?v=8qyHW2Mgc_Q&t=3m41s
So, UPST can offer lower APR to everyone on the entire FICO spectrum than its competitors, despite primarily lending to low-prime folks, because UPST’s superior AI underwriting is avoiding traditionally expected lower-prime default losses.
Savings from lower defaults are passed around to all.
Also, I dug up some more findings showing just how rigid other lenders’ underwriting models are. LendingClub and Prosper have been pricing unsecured loans with very traditional underwriting methods for years.
Here was some analysis by a PhD from May 2019 (I just looked him up, the author, coincidentally, currently works for Upstart in their machine learning department since being awarded his PhD in Finance):
https://uh-ir.tdl.org/bitstream/handle/10657/4628/CARMICHAEL…
“The mean FICO score of Lending Club and Prosper borrowers is in the 700–710 range, with the borrowers being tightly concentrated around this mean. Subprime borrowers are automatically rejected and borrowers with very high credit scores rarely apply…
…Overall, the summary statistics support the notion that the two lenders are direct competitors. They seem to be offering basically the same products to the same demographic.”
This is further evidence that at least LendingClub and Prosper have been doing nothing special vs UPST’s AI/ML. They are subscribing heavily to the idea of FICO score = ‘everything’.
Meanwhile, UPST’s senior VP of biz dev Jeff Keltner has stated in one of his podcasts that if UPST encountered a borrower who didn’t have a FICO score, their AI model would likely still function all the same, by using their other 1600 alternative data variables.