Citi’s 11th Annual FinTech Conference, Fireside Chat with Dave Girouard and Paul Gu.
My note paraphrasing this are below. I encourage UPST investors to still spend the time to watch the webcast in its entirety.
I thought the stuff said by Paul Gu was really good, regarding fraud and a depiction of how much more opportunity they have to improve their models which I believe would also improve their growth rates - gives me a sense that they don’t believe they’re running into a growth ceiling any time soon.
HOST: How is upstart auto retail software different from competing software offerings?
CEO: car buying experience is not pleasant for anyone. Dealers struggle with customer experience. Upstart’s Auto retail software is a frontend, think of an ipad application, that offers a smooth more pleasant and modern experience. We call it the Shopify of the auto dealership market. we’re also not just fixing the buying process but offering upstart AI loans so dealers can move more cars make more profit and consumers get lower cost loans.
HOST: Can you talk about how you’ve been driving adoption across dealerships? How’s the initiative going?
Do you see the potential to add other online market places, or even captive financing networks associated with OEM?
CEO: You’d think this is a small business selling process we’re going by email and over the phone and lightweight touch, generating leads and focusing on leads.
We focus on dealer groups with multiple rooftops.
A few months back it was people you could count on one hand selling this.
Since then we’ve invested a lot and now we had someone new to the team is already posing a five rooftop deal within a couple weeks of joining. it is definitely accelerating. We’d like it so a dealer could get started by downloading software, sign the agreement and not even have to talk to us. We’re not that far from that and moving quickly.
Everybody in auto lending ecosystem is potential partner of ours. They’re not competitors. Whether it’s captive financing networks, or even traditional banks currently in the industry, we’d like to work with them.
HOST: How should investors think about how this auto business scales in the next one to two years? How should we think of the unit economics for Upstart?
CEO: Anytime we’re entering a new space there’s a period of R+D and getting the ground in place. Then there’s an inflection point where models get better. This should be just like personal loans and repeat in auto but in a more compressed time frame. We expect it to become material to our business in 2022. Our teams are making fast progress on autorefinance and auto retail product. Unit economics should be similar to personal loan product. You can expect larger loan sizes than personal loans but percent take may be less. We feel confident our value add in this ecosystem is as good as it is in personal lending. so unit economics will not be dissimilar.
HOST: Where does UPST differentiate in machine learning and AI? How has the model evolved in last 3 years?
PAUL GU: Happy to be here and speak to that. My answer is also going to speak to why it’s difficult for other competitors to get started, not impossible, but challenging for others.
The core fundamentals of AI system is three things:
- Rows of data
- Columns of data
- learning algorithms.
You could think that you can just work on one thing and move onto the next and then get there.
But you realize that you need to do these three things in concert. If you only move one or two of these things you find out you need the other. For example if you have tons of training data but your algorithms are weak, you aren’t actually able to get any value from the effort of collecting the data.
Similarly if you have very powerful algorithms but your columns and rows are limited, you might actually have worse outcomes than the other way around as your models might overfit the poor data collected.
Institutionally it’s difficult for people to manage improvements in a similar pace in each three blocks which the blocks are already each difficult to gather, nd do them in the same time that none of them doesn’t bottleneck anything and still have proof points to the mothership to justify why this expensive effort should still continue.
What have we been doing for last 3 years? By 2015 and 2016 we established a foundation for this. We had a business and product that was essentially a ‘human specified rules based system wrapped around a series of ML models.’ The last 3 years we’ve iteratively gone to each part of that originally human designed system and said hey this thing that was a rules based system should be entirely replaced first with a model, and then replaced with a proprietary AI. Each place we had an assumption we replaced it with a model, and then replaced the model with a proprietary custom built AI.
As an example, Dave alluded to this on earnings call in the past. He said how it used to be the case that we had multiple ML models that were used in decisioning of loans. But it came out that the result came out of the models and they just got averaged together. Now increasingly starting a couple quarters ago we had the AI model outputs not be human specified logic but learning from the data directly. So that became an optimized weighted averaging similar to a traditional model. That was then thrown out and we said there should be no humans specifying the shape of how these models come together. They shouldn’t be a linear fashion in a human specified fashion. And so that structure should be a ML model itself. So we have multiple ML models feeding into multiple ML models. Each layer of this process is itself a proprietary AI built system. We still have a lot of work to do so not every corner of the code base is run intelligently, but it CAN be and it SHOULD be. We started from this core place where we had some ML, and now we branch outwads to have more and more abstraction of the whole system so we make the most intelligent decisions every step of the way.
HOST: So we start with portions of supervised AI and moving onto an unsupervised AI?
PAUL GU: In a certain way…now, those terms have a technical meaning in the ML world and don’t map perfectly to the intuition of what they mean.
Directionally what I’m describing is: we have one ML system and it produces an output and you have to decide what to do with it. Well a person has to decide what to do with that output. We want to replace that step. We want that step to be another AI system. And as you look across the entire product every place that you have a ML system producing an output and that output gets used, it’s an opportunity to have another ML system.
HOST: Can you help explain the interaction of alternative data with the models and how it contributes to better outcomes?
PAUL GU: So both the algorithms and data are entirely necessary. If you think of back to rows, columns, algorithms you need all of them and reasonable proportions to another or any of them are a limiting factor.
If you’re just stuck with limited set of traditional data like credit score and debt to income etc and throw a neural network in, it’s just not enough to exploit no matter how powerful the algorithm.
You need much more granular data in every category of data including the traditional data set. You also want orthogonal types of data that is lightly correlated with what you have. Like employment, educational history, how they interact with the application so that expands the total information space in depth and width. You can then pair it with learning algorithms that makes it work really well.
HOST: So 70% fully automated. IS there a limitation to how high the automation can go?
PAUL GU: The limit is one minus the percent of true fraudsters. We don’t want to give a fully automated loan to someone who is a bad actor. The good news about that is, except for short bursts of coordinated attacks, almost all of the time, true fraudsters are a tiny fraction of overall population of applicants. Most are trying to do the right thing to get a legitimate loan, sure some will round up on the income input but they’re not fundamentally misrepresenting who they are.
Our mission statement is true risk so that’s what it means here. True risk are the people who are true fraudsters. Our model gets a lot of people wrong so that’s opportunity. We look at a lot of metrics and it all says the same thing: only TEN PERCENT of all the kinds of error that exists in the personal lending system has been taken out by the work we’ve done over the past decade. The vast majority of it remains to be solved.
As you solve that, you can get the rate way up. You look at 70% and you think well how can that be when you’re already at 70%, but remember that people who get the automated experience are converting twice as much as those who don’t get fully automated. So there’s the rate at which you get automated fully at the front of the funnel as a lower number than in the back of the funnel so that means the opportunity is much bigger than what it seems from the 70% number.
[Recall on earnings calls and previous investor chats, they’ve all said that Upstart’s growth is dependent on innovation and model improvements…this statement here by Paul Gu is implying they are still far away from that ceiling - there’s lots and lots of runway for growth here.]
HOST: Last call you talked about developign differnt products. The value add is better credit decisioning. How can upstart compete in the prime category? does that impact the value proposition in contribution margin in a high versus low quality borrower?
CEO: We feel confident there’s a value proposition everywhere, even in the primest part of the segment… The best process is no process whatsoever… We’re also creating trust so the consumer comes back to us again… Opening to the entire credit spectrums it also means we can appeal to all Americans, we can market for example on national television when previously it wouldn’t make as much sense.
[Honestly not pleased with Girouard’s answer here, didn’t really give specific numbers to the question! He could have just said that he can’t say it at this time]
HOST: Small business product is a new territory for Upstart where there is entrenched competitors there. How can Upstart differentiate?
CEO: Small business lending is challenging as many have tried and failed to make it better. There’s different aspects to small business lending and we wont’ get involved with all of them, but for small business owners there’s a need for installment loans that are reasonable prices but super automated fast process the kind of things we’re good at doing. Making a product that appeals to business owners and still works for banks/credit unions is not a small challenge but is exactly the kind of thing Upstart is good at and something we can bring our DNA.