I listened to all available podcasts and youtube interviews with UPST management (CEO Dave Girouard, cofounder Paul Gu, senior VP business dev Jeff Keltner, product manager Alex Rouse, data scientist John Lewis, VP auto lending Val Gui). The biggest takeaway from watching all of these: UPST has extremely competent management, and they really know how to execute execute execute.
I further encourage everyone to watch the two interviews below. I highlighted some timestamps:
https://youtu.be/kRLzQ5b6en8
Ep11: interview with Sam Sidhu, CEO elect of Customers Bank
This whole interview is a must-watch. (The audience watching are potential bank partners.)
19:20 we had slowed down to 20-30M a month originations in middle of last year (due to covid), ended the year 50-75M a month, and now we’re glad to be doing 100M per month with Upstart
22:00 Upstart’s AI/ML for verification drives the verification and KYC costs down which makes lending much more economical and profitable. The instant approval simultaneously drives demand/conversion rates (increases value proposition to bank partners)
24:40 “why do you need 1600 variables? Isn’t it just FICO and one or two other variables “good enough?”
We found you need to get over 100 variables just to get to half of the explanatory power of our model…every little variable by itself is not super important even if the credit score itself is removed, many of the 1600 variables are saying similar things as they’re related…but it’s the ones that are related that are actually saying slightly different things and understanding how that reflects a difference in credit worthiness that gives you uplifts. To take advantage of this you need a combination of three things that no one really has done: rows of data, the 1600 columns of variables, and the algorithms. We know the truly creditworthy subprime borrower is there, when you realize when 20% of the pool defaults that sounds awful until you realize then that means 80% paid you back! It’s about finding the 80%. With $10 billion orginated we now have the hundreds of thousands of rows for the 1600 columns of variables. Now you can’t use your old logistic regression that goes to a score cord and prints on a 5 page PDF to take advantage. You need extremely sophisticated algorithms…and we even found some high credit score borrowers actually represent high degrees of risk…the sum of all those variables adding a little something together ends up making a really tremendous difference in understanding the creditworthiness of a specific consumer.
29:00 upstart increases value for bank partners via frictionless lending to a new customer which brings new crosselling opportunities for the bank (opportunities for repeat lending/opening bank deposits for the future)
32:40 in depth talk about upstart’s covid outperformance
44:35 we kept fraud below 30bps across history of entire platform (0.30%)
47:20 extremely important talk about Upstart’s effective fair lending. This is what banks want to see for compliance
53:10 upstart’s nps score of 80+ is massive versus traditional banks below 20.
https://youtu.be/nDmbkf5eUYM
Ep14, interview with Val Gui, VP Automotive Lending at Upstart
22:00 when we started auto refi the conversion rate wasn’t where we thought it would be, but once we worked with borrowers and got feedback on removing unanticipated friction and our instant approval increased to 50% (This is going from zero customers to 6 months later)
31:50 Upstart will be creating and leading the market of auto refi, (autorefi is currently less than 5% of entire autolending market) which the demand is there but the market has not yet been created to meet that demand
37:18 what’s a bold prediction you have for the future? **Oh man…there’s one but I don’t know if i can actually say it because it’s about Upstart…**I don’t know if as a public company I can say this and not get in trouble [just like with Paul Gu’s interview!]