Upstart CEO Dave Girouard Talks Machine Learning, Artificial Intelligence, and Personal Loans
Many or all of the companies featured provide compensation to LendEDU. These commissions are how we maintain our free service for consumers. Compensation, along with hours of in-depth editorial research, determines where & how companies appear on our site.
I recently had the chance to interview the CEO and Co-Founder of an interesting and successful company known as Upstart. Upstart made a name for itself over the last 6 years as a personal loan lender. How exactly did it do this?
Upstart implemented an artificial intelligence (AI) and machine learning (ML) algorithm that could underwrite credit applications without relying heavily on FICO scores.
In short, Upstart figured out a way to effectively lend to consumers who were being brushed aside by the traditional underwriting system despite being perfectly capable of paying back their loans. Furthermore, Upstart has even partnered with banks by offering its AI-ML algorithm and origination process in order to serve even more consumers.
Dave Girouard, the CEO and Co-Founder who spearheads this enterprise, hopped on a call with me to discuss the different aspects of Upstart’s business, underwriting, and the personal loans industry in general.
Read on to learn more!
Starting Up With Upstart
Q: Let’s start with a general background on how and when you got the inspiration for Upstart.
A: Sure, I was at Google for about 8 years and started the cloud application business there. At some point, I became interested in the notion of access to credit and how it worked. I began to form a belief that bringing modern technology and data science to credit could have a positive impact on who has access to credit and on what terms.
I observed that credit markets were somewhat antiquated, particularly around the credit decision itself which was the core. Much of the tech and many of the techniques we were using at Google under completely different circumstances could be applied to lending, allowing you to model risk much more accurately, thereby significantly reducing the cost of credit for a lot of people.
That was the heart of it. I was working in an unrelated business, but I became enamored with the idea of providing access to credit through technology and data science.
Q: What was it like transitioning away from Google to your own company?
A: It was definitely disconcerting. I went from a very big, successful company and running a very large team of more than a thousand people, to a team of exactly one. When I took the leap, it was a big deal.
Being an entrepreneur and starting my own company were not things I had done yet in my career. Even though it was a little nerve-racking and scary to start over, I was pretty energized and excited to do it.
Q: During the first couple years of existence, the startup stage, what was Upstart’s biggest challenge?
A: The biggest initial hurdle is getting the core team together. When you’re all on your own and have an idea, you might be excited, but you’re really just talking to yourself. It was challenging, but within a few months, I was able to assemble the core team, which was the first five or six people. So, it was a short-lived challenge that was ultimately very rewarding — six years later, they’re all still working at Upstart.
Machine Learning, Artificial Intelligence, and Credit Decisions
Q: Could you explain how exactly Upstart uses AI and ML to underwrite personal loans?
A: First, we gather as much data as possible about loan applicants. Far beyond traditional credit data, we look at things like education, employment history, how they interact with our application, etc. The amount of data in our credit decision is dramatically larger than what other lenders typically use.
Next, we use AI-ML techniques to help us understand the data we’ve gathered and to accurately model risk. These sophisticated AI-ML techniques are already successfully used in other industries (such as autonomous cars and smart home assistants like Alexa), but we’re the first to use them to make smarter lending decisions every day.
So, we’re using all of the data available and observing the performance of our personal loans on a day-to-day basis. Thus our model continually learns how to price the next loan. That’s the heart of ML.
Q: Could you elaborate on how your model is superior to the FICO model?
A: If a borrower has 20 or 30 years of credit history, then his or her FICO score will reflect whether or not he or she has made good decisions (paying back loans, etc). So, it can be reasonable in that case, but for many people, a FICO-based system doesn’t work very well at all. Younger people and immigrants who don’t have decades of credit history are generally penalized by it.
Our model is a superset of FICO-based or credit-based models. We actually have FICO and a lot of credit bureau data in our model, but this data provides a very limited understanding of the actual risk.
Since FICO is a very rough tool, adding more data gives you a better, more well-rounded view of the borrower. Here’s another way to look at it. FICO is a backward looking model; it’s a record of exactly what you’ve done with credit in the past. Whereas our model aims to signal your potential and what you’re likely to do in the future.
Q: Out of all this forward-looking data, what’s the most impactful on a credit approval-decline decision?
A: There isn’t a single variable that’s dominant in our model. There’s a lot of information, and any part of it contributes a small amount. So, there isn’t a single variable that tells us whether or not a person is high-risk or not.
This is really important to our core thesis. If there was just one variable, then it would be easy. Every lender would ask for that information, verify it, and model it. But the reality is the world is far more complicated than that; people and their lives are far more complicated.
If you want to have a decently predictive model, you need to look at a wide variety of variables and data about that individual and understand how they actually work together.
Q: Which piece of data surprised you the most? Or was there a certain revelation you got from all this data?
A: We found over time that FICO is not actually as predictive as most people think. There are plenty of people with FICOs in the 800s who will default on loans, and there are plenty of people with FICOs in the 500s who are actually very creditworthy. FICO is not useless, but at the same time, it’s far from being a dominant component of our model.
The Government’s Impact on Business
Q: How do the regulatory powers such as the CFPB feel about your AI-ML underwriting model in the lending industry?
A: Regulators are naturally curious and cautious about new technologies used in lending, but at the same time, they recognize that using alternative data and ML is actually a way to improve access to credit more than anything. After all, it can more readily identify a broader number of creditworthy individuals, and it can end up being a net positive for consumers.
We worked very closely with the CFPB in particular to basically demonstrate that our lending platform works within existing laws, particularly fair lending laws. Over a period of two years, we worked with them to demonstrate that appropriate safeguards could be applied to our lending techniques to ensure compliance with applicable laws. We were able to do that and were the first recipient of a no-action letter from the CFPB last fall.
Q: How has working with the CFPB impacted your business?
A: There’s certainly some overhead in highly regulated businesses like lending. We built a second system that effectively oversees and monitors how our lending platform is working and now report that data to the CFPB on a regular basis. That was a system we built and still maintain ourselves, so without question, there is some overhead in doing that.
However, it’s a reasonable price to pay considering you’re building a very dynamic and automated lending system that’s getting a little bit smarter every day. So, the quid pro quo is you need to monitor that system and make sure it stays within the bounds of the law and fairness.
Expectations: Grow, Profit, and Expand
Q: Can you share Upstart’s expectations for growth in 2018?
A: We grew originations by nearly 300% in 2017, and we expect to grow by triple digits in 2018. We’ve also achieved profitability in early 2018 and anticipate being profitable for the full year. So, the business is working really well in terms of high growth and profitability, and that’s a pretty unique position in this market.
Q: Upstart has a direct to consumer business as well as a bank partnership business. Which are you most excited about?
A: You know I can’t choose between my children. They’re both extremely valuable and important to us. We wouldn’t have a strong model with a lot of these advantages if we didn’t have a direct consumer lending program. But at the same time, to have the biggest impact possible on the market, we really believe you should get an Upstart loan from anywhere you need it. Banks are really in a position to take our technology much further than we could take it ourselves. That’s the rationale for making our technology available for banks.
Q: Do you plan on offering other products in the future?
A: Yes, I think the skills and technology that we’re developing will be applicable to many flavors of credit, and we definitely anticipate entering other types of unsecured credit. You can think of student loans and auto loans as in our roadmap, and someday even small business loans. Today, we’re focused on the consumer and unsecured consumer products. There’s a lot of great opportunity within that domain.
Upstart was founded in 2012 after Dave Girouard decided to leave Google to start his own venture. Upstart made a name for itself by developing an AI-ML algorithm for underwriting personal loans to applicants misrepresented by traditional FICO-centric models. Since its fruition, Upstart has originated over 140,000 loans amounting to $1.8 billion in funding.
About Dave Girouard
Before Upstart, Dave Girouard worked for Google, building the cloud apps business that everyone knows so well. Prior to his role as President of Enterprise at Google, Dave worked at Accenture, Apple, and Virage. Much of this experience helped inspire Dave to seek out a new underwriting credit solution, powered by AI and ML, as the backbone of Upstart’s success.
Author: Andrew Rombach