UPST CEO fireside chat with Deutsche Bank…

“What differentiates Upstart from all other lenders and competitors?”

There aren’t any others who really focus on improving the risk model surrounding credit origination. That’s the very center of what we do. There are others who work on the margins and very edges. We were first many years back now focused on disrupting credit decisioning itself. We have a better mousetrap. A product that can approve more borrowers at lower loss rates. That’s a magical thing. The thing about AI is it’s getting better as we get more data. Our models get more accurate as time goes by. That’s different than static, rule based models used by banks.

“Why aren’t there multiple competitors starting up now to compete with you guys?”

We wonder that ourselves. There’s a belief adoption curve. You first need a team capable of building machine learning models. Most ML engineers and founders don’t gravitate to lending because it’s so difficult to get into the business. It’s a rarity to have a team in this industry that can build these models and you need time and commitment for a system that needs to be built and refined over years. I’m quite sure others will if they haven’t started already as the opportunity in AI lending is so massive but we’re very happy to have a very significant headstart.

“When there are other competitors out there what differentiates you from them?”

Your data gathering and model tuning have to happen in concert. More variables and training data you get, the more your models can get sophisticated and support a more complex AI. But they have to grow in concert. You can’t short circuit it. A large bank with 30 years of data and past loans isn’t helpful because you didn’t collect the thousands of variables for each applicant. You don’t have the type of training data that we built over several years. There’s no way to short cut that process.

“What was impact of consumer lending demand by the pandemic and how is it changed since reopening?”

When COVID hit, banks pulled back in March 2020 including those on our platform as things were going crazy. Our volumes dropped precipitously. Unemployment went to 14%. Stimulus hadn’t kicked in. So people were seeking forbearance on their loans. But it was quickly clear to us our levels of forbearance request was dramatically less than the industry. There’s data you can see that [I’ve posted about this before] It became clear our model had a lot of advantages. One obvious explanation is the models consider what industry you work in. People in military, medical world were fine. If you were in travel, you were likely impacted. The models overperformed, it was really clear. Now stimulus kicked in, so first loan demand dropped. Eventually credit performance became crazy good. If you talk to anybody, default levels are very low. People aren’t spending, they’re paying off loans. We think the economy will naturally return to a normal state. We see our default rates are crazy low than predicted but we expect over time they will rise to what our models were initially predicting. [Actually in the latest KBRA presale report this past week, delinquency rates appear to have ticked upward although still remaining incredibly low]

“If credit tightens and defaults rise, what will be the direct impact?”
Our model expects that to happen. Banks are earning excess profits at the moment as default levels are below target. When they go back up, banks will go back to earning what they’re supposed to. We do think a lot of the industry is doing the inappropriate thing of passing along the lower default rates to the consumer in the form of lower prices. This is not a good thing because this environment is not going to be around forever. That’s the kind of sophistication and conservativism of how we do things at Upstart that matters a lot. In lending you can do things like lend as much as you want in a benign environment, it might work for a while. We have the strong system and discipline to not just go for the extra bit of growth.

“Conversion rates have expanded significantly. What improvement has caused this to increase?”

Conversion rate improvement happens for several reasons. As the model gets more accurate it identifies those likely to default and gets them out, as in not offering them loans, which increases the approval rate for everyone else. Then there’s the friction involved, what happens when you see a loan and you like it – automation has increased the conversion. The other dynamic is when the capital side gets more efficient in lower cost, as more banks go on the platform with depository capital, which is the cheapest form of capital, it means the cost of funding goes down and better offers go to consumers which increases conversion rate.

“This came up on your call where one bank excluded FICO scores. What would happen to conversion rates if you do exclude FICO scores?”

When a bank comes on our system they can apply a credit policy that overrides everything else. Our risk models operate behind that to price the loans. FICO been around for 30 years and banks use that as the fundamental tool to mitigate risk. What’s become clear is it’s really unnecessary. The model knows a lot more than a simple FICO score. What we start to see is banks go well, it’s 680 minimum, we’re not comfortable with 660. Then they see the model is working really well. They relax the requirements. They trust it’ll do the right thing. Ultimately we don’t care about FICO we really care about loss rates. It’s a great trend. That means the model learns more at the edges – certainly FICO isn’t useless, if someone has 550 FICO it’s for a reason, but it’s the nuance under the FICO score that our system goes to the source and understands all the details.
This will help our conversion rates and our model get stronger as less constraints are applied it.

“What’s the target conversion rate to hit?”

There’s an offsetting dynamic. If conversion rates get better that means at the top of the funnel we can market more effectively to more people. The marginal person you market to is less likely to convert. It’s not going to go up forever and could even go down, because we’re feeding growth at the top line. We don’t want to over-monetize. As a growth company with a long trajectory, we don’t want to maximize profits today, we want to gain market share quickly and expand the model because the more the model is fed with data the better it gets.

“What does the bank pipeline look like?”

The pipeline is very strong and it’s accelerating. I think having hundreds of banks on the platform, I feel very good about that in a few years, it’s not at all a stretch, we would expect that. But it’s important to say banks are very heavily regulated and a conservative industry so they don’t jump into something like AI without thinking very very hard about it. So it’s a long process getting banks comfortable. It matters the most when we become relevant to them in the categories they care about. Personal loans aren’t the center most important product for large commercial banks. But we’re now in auto lending, which is much more mainstream for large banks, and we’ll move beyond that. In our view, AI will be applied to almost all flavors of lending, and it’s only a matter of when. It’s a technology adoption curve not unlike cloud computing itself. When I started the cloud business at Google, in the early days it was a lot of small businesses and not a lot of fortune 500 companies at first. Fast forward to today its’ commonplace.

“Are those banks in the pipeline the larger banks or still small banks willing to test the waters?”

It’s certainly a lot more small banks and credit unions because they see peers doing it quickly. The largest banks on platform today is 40-50 billion. But now we’re moving up the food chain. You’ll see some banks I’m sure in the next year that are larger. Ultimately we think all the banks, including the big five, they’re going to need an answer, whether they want to build, buy, or partner. But we’re not in a rush, because bank adoption doesn’t gate our growth. We built a system that markets to consumers, banks take what makes sense to them, and then some banks flush the remaining to institutions in the capital markets. So the model keeps growing and getting better, and banks can get on board and participate when it makes sense for them. Bank adoption is not what’s gating our growth – bank adoption is about the long term strengthening of the ecosystem.

“How fast can you move into other types of lending?”

Every flavor of lending is different. We’ve been spending the last year and half moving from unsecured lending to secured lending in auto. We started with the most complicated, difficult underwriting of the individual without collateral. But auto is a different process, there’s a lien, title, perfecting lien. It’s taken time but now we’re beginning to wrap our auto business. But now Upstart’s learned a new trick. We now know how to deal with secured loans in addition to unsecured. So a different type of secured loan might be a home loan, similar to auto lending with similar challenges. We benefit enormously from what we’ve built, we’re not starting from scratch in new categories. I’m pretty confident we have the best and most accurate auto-finance models out there today, even our first versions of them, and they will only get better over time.

“Is the value proposition in auto or mortgage the same for banks/consumers as it is for personal lending?”

It’s basically the same value proposition. It’s an inefficient market. The vast majority pay too much. If you can get in there with a better mousetrap and frees up economics it becomes a better product for consumer and lender. In auto there’s a third party involved, the car dealer, and that’s an opportunity where we are confident we can deliver a better product for all three parties. That’s what a more accurate model can do. A lower priced loan for the consumer with better experience, a better loan portfolio for the bank, and the car dealer can sell more cars and make more profit. It’s going to happen quickly.

“How long will it take to get scale in auto?”

We’re putting all the pieces together this year. We have two parallel efforts. A refi product that’s up and running today at It was in one state at beginning of this year and 47 states now. Every state has different rules and regulations and we’ve refined that process. That product is scaling. In guidance it’s immaterial to 2021 but we’ve gotten the fundamentals in place. The other is getting footprint in car dealerships to have upstart loans offered at time of buying a car. We use Prodigy, a shopify for car dealerships. You can’t yet get an Upstart loan through it but that’ll change by the end of the year. We think 2022 will be a good year for growing and getting the auto product taking hold.

“What about international?”

It’s vast of course, in lots of parts of the world the credit system is far less functioning than the U.S. It’s wide open for someone like us to build a new model. We’re very excited, we have a solid base that’s growing quickly and profitable and gives us the change to launch a more initiatives. I expect over time getting outside the U.S. will be one of the initiatives.

“Looking at results, quadruple digit growth in 2Q, what’s causing this surprise?”

We’re conservative but not that much of sandbaggers, we’re surprised it’s exceeded our expectations, our AI models tend to improve with a steady organic rate of improvement we can almost predict. But once in a while upgrades to the system can all of a sudden increase conversions. Even though upgrades are in the pipeline we don’t always know the timing of them or the scale of impact. We’ve had new versions of models we thought would increase volumes a lot but they didn’t. Conversely this year, model updates did a lot more than anticipated. We have a lumpy growth business, where we have an organic rate we grow at but once in a while big upgrades cause step stair type of growth. It’s a little hard to predict. We’re aiming for long term strong position in this industry, it won’t be a perfectly smooth straight line, the technology matures as it does, even if we grow faster in some quarters than others as long as the models keep getting better we’re okay with that.

“Is there long term targets for revenue or adjusted margins?”

We’re in a category that the TAM is so enormous you couldn’t reasonably measure it, it’s in the trillions. The question is how much we can unlock. Our goal would be, we can continue to gain market share in the markets we’re in and unlock new TAM as we go. If we execute well, we can be a high growth company for a very very long time. TAM in the US and international. The question is how well can Upstart as a team execute and that’s something we feel really good about. There’s no world we see where we run out of runway or max things out. Even in our personal loan product there’s so much area for expansion and improvement. We now have auto, 6-7x larger market that we’re at the very first stage of. We’re not going to be serial as we have been to date, spending a long time to get personal loans right - you’ll see over time with Upstart, you will see us fan out over the next couple years initiate a few more products because we have now built the capacity to be successful doing that.

“What about M+A to supplement growth?”

We view M&A as a tool in the kit…we acquired Prodigy because we found a company that’s already built a product that could bring our loans to the dealership, we had thought it would take 2-3 years to build ourselves. It was a way to shorten the timeframe to get loans to the dealership. If M&A can get us somewhere faster, we’ll certainly look at that.

“Would you consider working with insurance companies so banks can share some of the risk?”

Openlending I’ve written about them before: [] is a company that does that. It pulls insurance company into the mix to restructure the risk of a loan. Our value proposition is different where we try to reduce the risk, not redistribute it. We have thought about it and we have relationships with insurance companies, it’s unclear but something we continue to look at.

“How about spending more marketing dollars to push consumers into the funnel?”

We spend all we can where the marginal dollar spent is profitable. We’re a methodical performance marketer. We spend tens of millions dollars a month to bring consumers to the platform. As conversion rates get better we dial up marketing. As we can serve larger parts of the US population we can begin to market in a broad fashion. We’re just starting to do TV trials on streaming TV. But we’re methodical. We don’t throw money into marketing as a nebulous concept, we are meticulous about measuring conversion and acquisition costs, etc.