Hi, this just got released this morning:
https://twitter.com/patrick_oshag/status/1443543934060404748…
My conversation with Upstart founder and CEO @davegirouard This is one of my favorite examples of an “AI first” company. Using models, they’ve improved the lending function for banks. Dave gives an incredible lesson on history and future of lending
https://www.joincolossus.com/episodes/91470262/girouard-maki…
My favorite quotes/highlights from the podcast by Dave Girouard. Brackets are my comments.
“And what our model very simply is trying to do is predict the cashflow for every month of a loan, which sounds simple, but actually it’s quite different than anything that’s been done in history, which is if you took out a three-year loan, there’s 36 monthly payments. What’s the chance that Patrick is going to either miss one of those 36 months and which one is it going to be or that he’s going to actually pay back that loan early during one of those months. So time-based cash flows is the essence of what our core part of our models predict and that leads to a pretty dramatically more accurate system.”
[A hint at UPST’s huge first mover advantage here. Time based cash flows means any new competitor needs TIME to train their AI models - on REAL-TIME loan repayment flows. Training an AI on past thirty years of a big bank’s loan data is effectively worthless.]
“…my co-founder at the time who was only 20, when we started the company, he would go out to a few websites and apply for loans and he would get rejected. And I was like, “Wow!” This guy went to Yale. He had a perfect SAT score. He was earning a six figure income. He had zero debt. He was getting rejected all around because he didn’t have three years of credit and I was like, “Wow!” I think one website gave him a 24% interest rate on a $10,000 loan and I was like, "wow!”
“…we’re trying to create a 360 degree view of the individual but we aren’t judgmental about the data in any way. We really are letting the software interpret performance. It’s not that Upstart thinks this school is better than that school or that this job is better than that job or anything else. It’s really the software interpreting and learning based on the performance of the loans…we actually test every single applicant for bias. We provide that data to the CFPB, the preeminent regulator for consumer protection, every quarter on behalf of all of our banks that work in our system.
So I guess maybe that’s a long-winded way of saying if you’re worried about AI bias the right answer is test all outcomes rigorously. We are set up such that if somehow our model moved in a way that it was biased against any particular demographic, we could revert, we could go backwards and avoid that. But fortunately what we’ve seen is our algorithms actually increase, improve credit outcomes for every single demographic we can name, race, gender, ethnic origin, et cetera, age, and that’s really powerful and regulators care about that. Higher approval rates, lower interest rates for every demographic and that’s a powerful combination.”
[Hard to not emphasize this enough, again and again. UPST has the ONLY CFPB No Action Letter for AI credit underwriting. Such a massive first mover regulatory advantage. Why should bank partners choose a future or existing new AI competitor over UPST, with this compliance peace of mind?]
“There’s so much sort of fear and loathing about AI and technology in general that to win that mind share and convince people AI can actually be an equalizer, is a really big challenge. We invest enormous amounts of resources in doing that every day.
Right now, my co-founder is probably talking to the CFPB multiple times a week and not just with the regulator, we’re talking about lawmakers, we’re talking about consumer advocate groups of different types. I’d certainly would love to have not had invested all that time and energy but at the same time, I think we’ve built a great moat. I mean, I think ultimately our business and our AI is very defensible in terms of it’s very pro-consumer. It is helping people get into the banking system that were not otherwise in the banking system and that’s ultimately our goal.”
[There you have it- that regulatory moat I keep saying over and over for months!]
“The thing I will say that is important is if you looked at the net interest income earned in lending in the US it is comparable to all profits in the technology industry. …the scale of it is so vast that people fail to appreciate how big the concept of lending is, but of course it’s why banks exist. So it’s just so vast that from our point of view, if you can actually make a significant improvement there not 10 basis points, but a hundred percentage points, the opportunity there is so vast and that’s why we’re focused on it. I keep going back to Google, but it reminds me in the early days of Google, Search was considered a commodity and all the sort of portals at the day we’re adding all these other features and just thought of Search as a check box. And then Google showed up and said, “No, actually that’s where all the money is.” And suddenly this company that was maniacally focused at that time on Search suddenly just won the day, and we think there’s something to that. “
“I mean, payments are obviously enormously important. There is so many giant companies being created in the payments sector, but lending is really where most of the profits are made. So it seems natural to us that there’s great opportunity there. Why are there more profits in lending than payments?
Well, I’d say because there’s more risks and unknowns. I mean, payments in some sense, tend to be somewhat commoditized in the sense of it’s a volume game. And if you look at the business models of Square or Stripe, it’s a very commoditized business, which of course means you want to add on other services around that. But it’s so vast of course, that you can build obviously giant, very valuable companies that are centered on payments, but lending is what just feels like this giant bespoke scary thing.”
“…my co-founder also says is 90% of the interest paid in this world is entirely unnecessary…We think of it from the lender perspective and it’s just such a crazy inefficient industry. Sometimes we think about Renaissance Technologies coming up with these trading algorithms, trading commodities, trading stocks, social bonds, et cetera, to kind of shave off THREE basis points. And we kind of think it’s laughable because we’re working at something that has two orders of magnitude more inefficiencies. So it’s that much of an opportunity.”
“…selling into banks is sort of a technology selling cycle in its worst possible form because banks are amongst the most regulated and conservative industries. We weren’t selling a chat app to put on your website or some sort of new thing for your data center. We’re selling something right at the core, an approach to lending, but we’ve built a team and we’ve said, we’re going to do this right, we’re going to prove it over time, we’re going to work, not just with banks, but with regulators. And it’s been a journey, but again, I think we’re getting there and the adoption rate, and it goes back to code computing or anything else that’s brand new. There’s a technology adoption curve and you have to both respect it and then move it along and as quickly as you can.”
[UPST has been able to grow so fast because it has gained the trust in the capital markets with whole loan sales and securitized trusts to quickly expand loan growth. They would be nothing if they only decided to simply bank partner (as in banks who largely only retain loans on their balance sheet). The bank partner ecosystem is a long term slow growth strategy. It isn’t going to be quick. This is interesting to me - UPST would otherwise be a very very tiny company struggling to breakthrough and scale, like its competitor ZestAI (which was founded even before UPST), if UPST management didn’t choose to rely early on this capital markets path. Very insightful to me about how the leaders of the company is just so so incredibly important. The quality management team at UPST is really a diamond in the rough.]
It’s always nice and valuable to listen to these interviews/calls. Dave — as always — exudes confidence in what the team does and what they are working on. Whenever I listen to these calls, I feel even more confident with my overweight allocation. Here are my top 5 Q&As along with my comments.
Q: [00:02:54] The topic to begin with is this idea of speed as a habit in business. And this is very popular notion in technology, the move fast and break things, probably the most famous example of a phrase that exempt flies this idea. But I think you’ve really thought about the various levels of building speed into a habit and why that’s good and maybe ways in which it isn’t good, or it can be dangerous, and I’d love to just begin with this concept.
A: [00:03:35] Basically I came down to the idea that most of what we do in business is number one, making decisions and number two, acting on those decisions and then really as a sort of third thing, a lot of times you have to sort of get others, other companies, other people on board with those decisions. And if you could do all three of those things faster habitually, you’re going to just have more trips to the plate, more swings, whatever you want to describe it as. And that was the idea that went down to how a company makes decisions. I was just really into the effort of how you make decisions and which ones should take a long time or short time and just how you get things done.
My belief was, particularly for founders, but really for any type of executive, is you can begin to reinforce behaviours that will make your company move more quickly. And that was the heart of that. It was really thinking about, and maybe one of the most foundational ideas is the most important thing about a decision is to decide upfront how long you’re going to take to make that decision and who needs to be involved in it, which sounds like an obvious thing, but most people don’t really think that way. And if you do that upfront, you’ll find you may not always make the right decision, but you make an expedient decision and that’s almost always better.
[My thoughts: I really like the fact that Dave as the CEO wants to move things as fast as possible. He wants to make speed a habit as he calls it — especially for founders/executives —by reinforcing get-things-done-fast behaviours. And by focusing on how much time you will spend on a decision upfront you avoid wasting time which is even better than waiting for the “right” decision. This tells me that if there is a way to roll out new products he and the team will do so as fast as possible. Their goals are aligned with ours. As investors we want them to move into other verticals as soon as possible.]
Q: [00:05:38] So if there’s room for improvement or faster speed at three levels, making the decisions, and maybe there’s even one before that, which is knowing which decisions to spend any time on, but making the decisions once you know you have to make one and executing on them and sort of convincing or getting others that matter outside the business to move at the same speed. I’d love to just pick apart each of those, I’ll call it four stages, and just hear what you’ve learned. So I guess the first one is how do you know that something is deserving of a decision in the first place? That seems like a triaging that has to happen before we get to the other three. What have you learned just in selecting your decisions in the first place?
A: [00:06:12] As a CEO of a company your goal is not to make more decisions than you need to. You want to build a company capable of operating where important things are done far below you in the org.> And so that’s the first thing is, if I’m involved in every decision the company is making, that’s a bad sign. I should be breaking ties or helping to work some of the most important decisions that I have a unique perspective on. So that’s the first thing. And then beyond that, it really comes down to, we’re all making decisions all day long. We just have to discern which ones are worth two minutes and which ones are worth two days and who should have input? And the trick as a leader, I think generally is to give people a voice but don’t try to create consensus because consensus is almost impossible if the decisions are even, we’re thinking about.
[My thoughts: I like the fact that he doesn’t want to be involved in every single minor thing in the organization by micromanaging every single person. Micromanagers usually have the exact opposite result of what they initially thought. Dave doesn’t just want to appear as a forward-thinking leader. But he actually wants to be one. This was also evident when he made the digital-first announcement in the earnings call a month or so ago.]
Q: [00:06:59] So group decision-making doesn’t exactly work well. So if the first thing is making sure you’re not the bottleneck and the second is sort of identifying how much time you’re going to spend on a given decision. What have you learned about the execution layer of all this? Once you’ve made a decision, how does speed then translate into what the teams actually do?
A: [00:07:31] Execution often starts with meetings and discussions as well. As much as we all hate the concept of meetings and people love to bash meetings, it frankly is how a lot of things get done. But in meetings you have to organize a plan and execute a plan and oftentimes there’s not enough consideration to time. How many times I’ve sat in meetings where somebody would say, “Okay, here’s the next two things or three things we’re going to do. Thanks. Let’s go.” And no one ever said, “When are we going to be done with those? How fast can we get that done so we can move on to the next thing?” And having a habit of asking, when will this be done? Why can’t it be done sooner?
[My thoughts: The CEO is laser-focused on getting things done. I remember in an early interview of his when he described himself as a results-oriented CEO and not an Elon Musk kind of CEO. If you’re looking for a crazy, visionary founder, I don’t think anyone describes me that way. I look like the CEO you bring in when the founder goes off the rails." You can read that interview here (if you haven’t already) review.firstround.com/fresh-off-ipo-upstarts-ceo-shares-why-…
Q: [00:18:40] Can you describe credit score and I’ve realized I’ve never really thought about credit score before. What drives it? Why is it suboptimal as the sort of predictor of why someone will pay or won’t pay repay a loan? And then I really want to hear the full story of building an AI function like this or a data model function like this into a core product for business, because I suspect we’ll see this applied in a million other industries over the next couple of decades.
A: [00:19:16] So the credit score was invented about 30 years ago, 1989 ish and before that, I mean, there was nothing. If you were going to try to get a loan, you would sit down across the table from somebody, a loan officer at a bank, and they’d ask you a bunch of questions, they’d have a bunch of rules, they’d want to know where you work and what you paid for your mortgage, et cetera. And that obviously was a very bespoke process with all sorts of problems, it had problems of fairness, problems of accuracy, of performance. And when Fair, Isaac came out with the FICO score in 1989, suddenly there was this universal three digit number that gave a sense of how credit worthy you are. So at the time it was a radical leap forward for any particular type of bank or lender trying to make a credit decision to actually have a number that means something based on your prior use of credit.
30 years later, it still is the centerpiece of how credit decisions for consumers are made, whether that’s a credit card, a mortgage, the government uses it to decide which mortgages can be sold to the government sponsored entities, Fannie Mae and Freddie Mac, and it just became encrusted into how the world of credit works. But when you think about it for a moment, a three digit number is never going to capture all the subtlety of a person and whether they would pay back a loan and when, and what size of loan, on what type of loan. So we had a huge pallet of opportunity to improve on that.
[My thoughts: If 30 years ago FICO score was a radical leap forward, then what Upstart is doing now is surely logical and the natural move of things considering what technology we have available. As the first mover, Upstart has a tremendous opportunity to change this forever (or at least for the next 30 years or so). Speed is critical here too as they want to capture as much opportunity as possible.]
Q: [00:33:07] Is it fair to say that if I think about the compounding asset that you are trying to foster or build at Upstart is this core data model and data set that you’re just every year getting better than the next relevant competitor at knowing something about the consumer in question here and that with that knowledge you become almost like a platform. If you’ve got that insight on a customer, that insight can be used right now to make loans but potentially for lots of other financial services too. So do I have that right? And if it is right, how do you think about intentionally compounding that value because I assume that’s where a lot of your competitive advantage comes from?
A: [00:33:47] They are centralized artificial intelligence models. So every consumer that gets a loan, every bank that works with us is both contributing to and benefiting from these centralized AI models. So we do get this advantage because the model with more data, more experience is the best one. Now the other part is we have to teach it new tricks. We started with unsecured lending, which is a very simple form of credit. Now we’ve moved into auto lending, which is a secured form of credit, right? You have the consumer to underwrite, but you’re also going to have this asset that’s backing the loan, which in this case is the car. So now we’ve sort of taught the model new tricks and instead of being backed by an automobile, it might be a home. It might be the cash flows of a small business. It might be a piece of heavy equipment.
So our idea is generally centralized models that are learning as quickly as possible and then taking sidesteps to learn new tricks where suddenly almost any type of lending in the world and potentially domains beyond lending, we have significant advantages in. And I think that’s the heart of what we’re building is an AI model that can properly price almost any flavor of credit, it just gets a little bit better at doing that every month, every week, every month and that’s enormous opportunity. Not easy to navigate, a lot of things you could do wrong, but I think with strong execution it’s almost unlimited addressable opportunity there.
[My thoughts: The thing I like about Upstart is that they are a technology company solely. They do not underwrite the loan or carry it in their books they simply make it easier and fairer to acquire. This has almost no limitations compared to let’s say Affirm that needs the funds to back it up. Also, by teaching their models “new tricks” they essentially prepare them to be ready to be applied to a new flavor of credit. Just like compound interest, the models learn more and more as time goes by and as more tricks are thrown into the mix. Secured, unsecured, whatever it might be. So all these decisions today, help to form the AI model of tomorrow which can work almost on any flavor of credit. That is an enormous advantage over potential competitors who will need both fresh data and adequate time for their models to learn and improve. There’s simply no shortcut.]
Pavlos, I really liked your post. Thanks. I found your thoughts which you take on to be very useful. For example:
[My thoughts: The thing I like about Upstart is that they are a technology company solely. They do not underwrite the loan or carry it in their books. They simply make it easier and fairer to acquire. This has almost no limitations compared to, let’s say, Affirm, that needs the funds to back it up.
Bolding was mine. That’s exactly what I’ve worried about with regards to Affirm. It seems to me to be very cash intensive. Gaucho Rico wrote that they had $2 billion in loans on their books at the end of June, up from $1 million a year ago. Well they had an IPO to raise cash. What will they do this year when Walmart, Amazon and Shopify come on full bore? Sell more stock to raise the cash they will need? I know that they are trying to securitize those loans and get someone else to take some of them off their hands. But the way their business works (as I understand it), they have to pay the merchant for the merchandise up front (at a discount, which is Affirm’s fee), and collect over time from the individuals who are buying now and paying later. Boy that sounds capital intensive. If I take a position in Affirm it will be very small.
I also really liked this observation about Upstart:
Also, by teaching their models “new tricks” they essentially prepare them to be ready to be applied to a new flavor of credit. Just like compound interest, the models learn more and more as time goes by and as more tricks are thrown into the mix. Secured, unsecured, whatever it might be. So all these decisions today, help to form the AI model of tomorrow which can work almost on any flavor of credit. That is an enormous advantage over potential competitors who will need both fresh data and adequate time for their models to learn and improve. There’s simply no shortcut.]
Bolding was mine. That’s exactly what I’ve worried about with regards to Affirm. It seems to me to be very cash intensive. Gaucho Rico wrote that they had $2 billion in loans on their books at the end of June, up from $1 million a year ago. Well they had an IPO to raise cash. What will they do this year when Walmart, Amazon and Shopify come on full bore? Sell more stock to raise the cash they will need? I know that they are trying to securitize those loans and get someone else to take some of them off their hands. But the way their business works (as I understand it), they have to pay the merchant for the merchandise up front (at a discount, which is Affirm’s fee), and collect over time from the individuals who are buying now and paying later. Boy that sounds capital intensive. If I take a position in Affirm it will be very small.
Saul, They keep loans on their books, but as you noted they also securitize and resell it to other institutional investors. And securitization will be progressively more important going forward.
Here is what their CEO had to say on securitization:
BRIAN CHEUNG: Hey, Max. It’s Brian Cheung here. So let’s talk about your company’s payment structure here. So my understanding is that you used to rely on warehouse lending to support what is essentially a loan through this buy now, pay later program and other products that you have.
Now you’re doing a little bit more securitizations. Wouldn’t it be easier just to have a bank charter and be able to use the deposits on your balance sheet to do all of this? Or is the challenges that comes with crypto, if you’re a traditional bank, kind of something you want to do, which is the reason why you haven’t gone down that path?
MAX LEVCHIN: So first of all, you’re exactly right about our funding strategy. We started out literally funding out of equity where no one knew what this thing would be like and would it work. And so we put up our own money to fund these transactions. Over time, we got more and more market confidence, expanded into warehouse lines. We now securitize, and the market has given us excellent reception.
I think we think about all these things as enabling factors for product. Fundamentally, we’re a product and engineering-led company. That means that when we think of things, you know, what do we do next, it’s always in the context of what do we do for our partners, the merchants, and what do we do to delight consumers?
Source: https://news.yahoo.com/50-population-plans-buy-now-151043559…
Apparently demand for their loans is high.
Also for loan originations, they use they get it from Cross River Bank and Celtic bank.
Affirm’s business model is more complex than that of many Main Street banks, which use deposits on their balance sheets to fund loans. Affirm relies on two banks—Cross River Bank in Fort Lee, N.J., and Celtic Bank in Salt Lake City—to originate most of its loans. It then buys the loans from its bank partners and services them throughout the life of the loan…Investor demand so far for Affirm’s securitizations has been strong, said Imran Ansari, a senior vice president at DBRS Morningstar, adding that the company’s funding model provides it with other options should that change. “If any one funnel shuts off, they have another outlet,” Mr. Ansari said.
Source: https://www.wsj.com/articles/fintech-lender-affirm-leans-on-…
So I am likely not too worried about their ability to get capital. Credit risk is though higher than Upstart with Affirm’s model. It is more Amex than master or visa.
In the end the eventual winner in this space will be the one who can price the risk most effectively.
I wonder if there are industries beyond Lending that $UPST can disrupt?
Lending isn’t the only thing “Credit History” is used for; it is are also used in:
*) Employment applications
*) Background checks e.g. for certain Government positions requiring Clearance / Elevated Trust
*) Rental applications
Saul, here is my understanding of how to assess Affirm’s capital efficiency.
Let’s see their definitions of 3 important KPIs first:
Total Platform Portfolio - The Company defines total platform portfolio as the unpaid principal balance outstanding of all loans facilitated through its platform as of the balance sheet date, including loans held for investment, loans held for sale, and loans owned by third-parties. The Company believes that total platform portfolio is a useful financial measure to both the Company and investors in assessing the scale of funding requirements for the Company’s network.
Equity Capital Required - The Company defines equity capital required as the sum of the balance of loans held for investment and loans held for sale, less the balance of funding debt and notes issued by securitization trusts as of the balance sheet date. The Company believes that equity capital required is a useful financial measure to both the Company and investors in assessing the amount of the Company’s total platform portfolio that the Company funds with its own equity capital.
Equity Capital Required as a Percentage of Total Platform Portfolio - The Company defines equity capital required as a percentage of total platform portfolio as equity capital required, as defined above, as a percentage of total platform portfolio, as defined above. The Company believes that equity capital required as a percentage of total platform portfolio is a useful financial measure to both the Company and investors in assessing the proportion of outstanding loans on the Company’s platform that are funded by the Company’s own equity capital.
So “Equity Capital Required” is the real capital that “the Company funds with its own equity capital.”, not the 2B “Loans held for investment” in the balance sheet ending Jun 30, 2021. And “Equity Capital Required as a Percentage of Total Platform Portfolio” gives us an even better picture of how Affirm improves its capital efficiency.
Then let’s see how these metrics improved in the last couple of quarters since IPO.
Equity Capital Required (m)
2020 229.5 220.8 220.4 277.3
2021 206.6 178.1
So in an absolute number basis, “Equity Capital Required” was trending down to only 178.1m in Q2 (down 19% from a year ago!)
Equity Capital Required as a % of Total Platform Portfolio
2020 9.8% 8.9% 7.6% 7.5%
2021 4.9% 3.8%
And “Equity Capital Required as a % of Total Platform Portfolio” was even more impressive to now only 3.8%, hugely down from 8.9% a year ago!
I think Affirm management did a fantastic job in managing their capital and now with 178.1m their own capital and only 3.8% against Total Platform Portfolio, I don’t think they are as “capital intensive” as most people think.
Zoro
And what our model very simply is trying to do is predict the cashflow for every month of a loan, … nd which one is it going to be or that he’s going to actually pay back that loan early during one of those months
This is a really interesting point that I hadn’t really thought about with regards to UPST. I wonder how much their models are trained to not only find borrowers that are credit-worthy, but also borrowers that are profitable.
There’s this irony that happens after you make a loan: you receive money from the people that you don’t want money from, and don’t receive money from the people that you do want to receive money from.
If UPST is effective at making decisions not just on who is the most credit worthy, but also the entire profitability picture and cash flow of that loan, that is an interesting differentiator.
–CH
(long UPST)