Upstart and Capital One Auto Loan Navigator

So, one of the key reasons I’m excited about Upstart is its nascent auto loan business. Why?

  • The addressable market is much bigger than personal loans. As noted in their investor presentations.
  • I like collateralized loans better. Uncollateralized loans are obviously risker and that means their value fluctuates more. Yes, it’s not Upstart that bears that risk directly, but it still concerns me from a macro level.
  • I believe auto loans are more open for disruption. Uncollateralized loans, except for the risk, are easy. To originate a loan, it’s just you and the borrower. Upstart becomes a third party, of course, but they only play a simple (if powerful) role as the arbiter of risk. Auto loans are much more complicated. The government is generally involved, because you have a car title. (And that process varies widely based on state and local government.) Plus an auto dealer. Plus, the regulations are often much more stringent around auto purchases and loans. Plus, there is a whole other element to the decision: the car and the car’s depreciation schedule.

With this background, I’d like to introduce you to a technical presentation from the Capital One Auto Navigator team in 2017.

This is a technical presentation, and 90% of the technology talked about in this presentation is not relevant to Upstart, so I want to just pick out a few parts to highlight. Both to translate some technical things, but also to add some timestamps for a few interesting business nuggets.

The presentation: (I’m not going to create YouTube timestamp links because of how TMF links work. You’ll have to jump around yourself.)

  • 01:06 , Explains that Capital One is the #1 auto lender in the US. (He doesn’t explicitly say this here, but he means outside of the captive loans from the auto manufacturers.) Then he explains a bit about the pains of the traditional process for borrowers.
  • 03:00. , Nice visual on the high level flow of a loan decision. (Note that Cap1 is still very FICO dependent. At least they were in 2017.) But he highlights that they need to run 1000’s of ML models for each application.
  • 19:30 , A bit about the performance characteristics of their system. Mostly technical, but the key point I really wanted to emphasize is that they spend the majority of the processing time waiting for the traditional credit bureaus to respond.
  • 22:21 , The TV ad for Auto Navigator, just to explain what Capital One thinks the value to customers is.
  • 23:30 , A tiny bit on their use of Big Data to handle the millions of cars in the car data. I include it because it highlights that auto lending has additional data sources to think about.
  • 32:46 , A bit on how they take new ML models from the business and how they have to continually update them to react to the market. And specifically how they run up to 20 ML models in parallel, and have guardrails around those models.
  • 34:35 , A bit on auditing. Illustrating how the regulations affect their ML: the government can ask them up to seven years later to justify their loan decisions. (Side note for Confluent fans, they use Kafka for this.)
  • 40:41 , A bit on how they develop the ML models.

If you don’t want to listen to it that’s fine because the key points I really want you to take away are:

  • Using ML for loan decisioning is not new. Capital One was already live, in production, with a pretty sophisticated system 4 years ago.
  • Despite that, and how big the car loan market is, there still aren’t many players that try to compete at the dealership. Search for “car loan iPhone app” or something similar. You will likely get results that include Capital One, refinancing (much easier), simply payment calculators, and maybe an aggregator or two (like LendingTree). Dealerships are hard to penetrate.
  • Banks are still interested in making auto loans though. Search for “car loan” and you will get a much broader set of results. But if you follow the links the banks tend to link you out to partner websites very quickly.

My thinking?

  • Do I believe that Upstart is unique in offering ML based loan decisioning? No.
  • Do I believe Upstart has the best ML? Maybe? The superb retail banks are quite good an ML, and they do have a lot of data to work with.
  • Do I believe that the traditional banks are focused moving past FICO the way Upstart is? I definitely could be wrong here, but I don’t think so. The big banks are too conservative and the small banks don’t have the resources. The banks I work with are more focused on how to use ML to cross-sell/up-sell than on how to use ML to complete aggressively on rates. [But again, I could be wrong, not my current tech focus.]
  • Do I believe that there is an huge segment of banks that aren’t big enough to build their own that will want an off-the-shelf solution? Absolutely. And the market is huge.
  • Do I worry about Capital One as a competitor? Absolutely not. To me this is just market validation, especially since Capital One is known as a bit of an innovator. If I’m a small bank, do I want to do all of the complicated tech outlined on this presentation on my own, including keeping my models constantly refreshed? Hell no. Even as a regional bank, if Upstart can prove to me that they will make industry leading decisions, I’d be happy to outsource it to Upstart.
  • Do I believe that the customer experience still needs a huge leap forward of innovation? Absolutely. A lot of other auto loan sites feel pretty scammy to me.


I really worry about Upstart. Not just because of the wild ride in the last few months. But because it’s a non-subscription business that is competing with big banks in what they think is a core competency.

But, despite those reservations, I remain confident in the company (and stock) overall. Because I feel like they are just hitting some of the network effects that we love so much as investors. Both from a technical level, but also from a reputation/customer level. My experience in the banking business is that small, medium, and even regional banks are extremely happy to use solutions like this. They just don’t like to be early: they want something that is well proven and widespread. Especially for something they consider a core business.

With today’s dip, I just couldn’t resist buying a little more. I will remain vigilant (and somewhat unforgiving) as I do with all of my “Saul” stocks, but I feel like if they can gain a toehold in auto loans the upside is tremendous.



Apologies for the short post, but, as always I find a few typos in my original post. (And TMF has no way to let you correct them.) Usually they are pretty harmless, but I think this one potentially is confusing.

But I see that my browser has autocorrected my statement “The superlarge retail banks are quite good an ML” to “The superb retail banks are quite good an ML”. While I don’t have anything against retail banking, I didn’t mean to imply that all retail banks are superb. :slight_smile:

The point was meant to be that while the major retail banks (Cap1, BofA, JMPC, etc) are quite good at ML, I can’t say the same about credit unions, and small/medium banks. Regionals staddle the difference.

Since I find myself making a short post, let me highlight Fiserv as an example of how financial services companies (especially small ones) are willing to outsource core parts of their business once issues of vendor viability and vendor risk are mitigated. Fiserv’s business model is practically "consolidate fintech companies, including core banking and card processing, and then market them (primarily to small/medium financial companies, but not exclusively) to a larger audience based on the stability of Fiserv.

That’s not to say that I think Upstart is going to get acquired by Fiserv. Upstart is probably too hot right now for that. I’m just making the point that once you gain critical mass in the financial services industry, there are network effects.