Menu Close

Upstart Holdings: Victory for AI? Or Not…

This is a syndicated repost published with the permission of The Institutional Risk Analyst. To view original, click here. Opinions herein are not those of the Wall Street Examiner or Lee Adler. Reposting does not imply endorsement. The information presented is for educational or entertainment purposes and is not individual investment advice.

June 17, 2021 | In this latest edition of the Premium Service of The Institutional Risk Analyst, we ponder the nouvelle cuisine known as artificial intelligence (AI) applied to the world of unsecured consumer lending. Upstart Holdings (NASDAQ:UPST), a portfolio company of Third Point LLC, is one of the more prominent and successful examples of machine learning applied to creating and managing unsecured credit risk. Will it work “through the cycle?” Ask us in 2025.

As we told the FDIC and other regulators in our comment letter on AI as employed by insured depository institutions:

“While the progress in machine learning is impressive, horsepower in terms of computational capacity does not translate into intelligence. In fact, machines, are never going to be ‘intelligent’ in the way that we apply this term to human beings. As a result, prudential bank regulatory agencies as a group must parse the difference between the promise of AI, as marketing and media concepts, and the technical and practical realities as applied to owning and managing insured depositories.”

Former Warburg Pincus Vice Chairman William Janeway noted in an October 2018 interview in The IRA (“The Interview: William Janeway on Capitalism and the Innovation Economy”):

“AI systems seem to be good at pattern recognition when they have been properly trained as to the pattern in question. They are good at playing games where the rules of the game are given exogenously such as in chess or go. They are good at that. But the games that really matter, like the Three Player Game, are those where we must co-invent the rules as we go along. For example, in any conversation, even with people you know well, you are constantly trying to understand the context of the words used by the other speaker and vice versa.”

What Janeway illustrates is that no matter how much you train an AI system to identify patterns, to apply quantitative methods to understand human behavior, the ability of consumers and creditors to change the rules of the game as they go through time renders such systems vulnerable to failure or deliberate gaming. But beyond the question of the fragility and brittleness of AI systems, however, there is the broader issue of how AI is being used today to replace or augment people involved in managing credit and operational risk.

Upstart as Challenger

Companies like UPST represent a challenge to traditional credit risk management. The world of FICO scores and collateral are the twin pillars of secured consumer finance, but unsecured lending is entirely another matter. From the relatively pedestrian unsecured consumer loan portfolios of large banks such as JPMorgan (NYSE:JPM) and Bank of America (NYSE:BAC), to more risky consumer lenders like Citigroup (NYSE:C) and Capital One (NYSE:COF) with default rates measured in points, to subprime lenders with gross loss rates that stretch into double digits, unsecured consumer credit is a profitable but also a hazardous business.

Over the past five years, credit data providers led by Experian (LON:EXPN) have sponsored a challenger to Fair Isaac (NYSE:FICO) and the FICO score, Vantage Score, which uses non-traditional measures such as utility bills and social media profiles to make a determination about the borrower’s ability to support credit. UPST represents an operating business built upon such ersatz metrics, combined with intensive data mining of borrower behavior. UPST describes itself in its 2020 10-K:

“We are a leading, cloud-based AI lending platform. AI lending enables a superior loan product with improved economics that can be shared between consumers and lenders. Our platform aggregates consumer demand for high-quality loans and connects it to our network of Upstart AI-enabled bank partners. Consumers on our platform benefit from higher approval rates, lower interest rates, and a highly automated, efficient, all-digital experience. Our bank partners benefit from access to new customers, lower fraud and loss rates, and increased automation throughout the lending process.”

Translated into plain language, UPST is essentially a lead generation platform that finds borrowers, presents these borrowers to banks for funding and sometimes retention in portfolio, and sells some of the production that the banks don’t want to investors. UPST is kind of a mix between a wholesale channel for generating leads for unsecured consumer loans and the marketplace concept of Lending Club (NYSE:LC), including a securitization channel. The market value of the common equity of UPST is up over 300% in the LTM. Some of the obvious comps are shown below:

UPST describes its business:

“Loans issued through our platform can be retained by our originating bank partners, distributed to our broad base of institutional investors and buyers that invest in Upstart-powered loans or funded by Upstart’s balance sheet. In the year ended December 31, 2020, 21% of the loans funded through our platform were retained by the originating bank and 77% of loans were purchased by institutional investors through our loan funding programs. Our institutional investors and buyers that participate in our loan funding programs invest in Upstart-powered loans through whole loan purchases, purchases of pass-through certificates and investments in asset-backed securitizations.”

Of note, UPST discloses that it does not own a broker-dealer despite its sales of loans and ABS to investors. This is key business model difference with SOFI, for example, which uses a FINRA regulator broker dealer as its platform. Lending Club has become an FDIC-insured bank. In this regard, UPST provides a key piece of information:

“A large fraction of the whole loans sold to institutional investors under our loan funding programs are originated by Cross River Bank, or CRB. In the year ended December 31, 2020, CRB originated 67% of the loans facilitated on our platform and fees received from CRB accounted for 63% of our total revenue. Our current agreement with CRB began on January 1, 2019 and has an initial four-year term, with a renewal term for an additional two years following the initial four year term. We enter into nonexclusive agreements with our whole loan purchasers and each of the grantor trust entities in our asset-backed securitizations, pursuant to which we provide loan servicing.”

How does UPST manage to originate loans that other firms cannot or will not? Via artificial intelligence and using non-standard credit metrics other than FICO scores. While these data points are new and innovative, they are not tested in a down credit environment. UPST attempts to address these concerns:

“Credit is a cornerstone of the U.S. economy, and access to affordable credit
is central to unlocking upward mobility and opportunity. The FICO score was invented in 1989 and remains the standard for determining who is approved for credit and at what interest rate. While FICO is rarely the only input in a lending decision, most banks use simple rules-based systems that consider only a limited number of variables. Unfortunately, because legacy credit systems fail to properly identify and quantify risk, millions of creditworthy individuals are left out of the system, and millions more pay too much to borrow money.”

Translated into plain language, UPST lends to borrowers who cannot access credit via a FICO score based credit process. These borrowers may not even have sufficient utilization of credit to generate a FICO score. UPST continues:

“We leverage the power of AI to more accurately quantify the true risk of a loan. Our AI models have been continuously upgraded, trained and refined for more than eight years. We have discrete AI models that target fee optimization, income fraud, acquisition targeting, loan stacking, prepayment prediction, identity fraud and time-delimited default prediction. Our models incorporate more than 1,000 variables and benefit from a rapidly growing training dataset that currently contains more than 10.5 million repayment events. The network effects generated by our constantly improving AI models provide a significant competitive advantage—more training data leads to higher approval rates and lower interest rates at the same loss rate.”

So, the basic thrust of the UPST model is to use AI to originate and sell consumer loans, whether to banks or to investors a la LC, SOFI, et al, but without the cost of being a bank or broker-dealer. This originate-to-sell model also includes less kind comparisons including Citibank (2008) and Greensill Capital (2020), the latter which originated commercial paper for sale to investors, but collapsed amid allegations of fraud. Third Point, however, claims that AI has helped UPST deliver truly exceptional credit results:

“Upstart’s AI platform has yielded a 75% reduction in loss rates. Upstart’s ongoing model improvements deepen its competitive moat and continually strengthen its business case. This is reflected in increased credit rate requests and increased loan conversion leading to ~87% CAGR in the total number of loans transacted. Upstart’s model benefits from flywheel dynamics that should drive compounding growth through a cycle of continuous model feedback and improvement. As the platform grows, more data points (payments, defaults, etc.) are fed into the model, thus improving its accuracy and supporting additional share gain. An understanding of the opportunities presented by AI and machine learning has been an important theme we have expressed in numerous investments at the firm.”

Of course, terms such as “moat” and “flywheel” are not defined in GAAP, but you get the idea. It’s different this time, we’re “going to the moon” a la Reddit. As we noted in our comments to the FDIC on AI, one of the difficulties in assessing UPST and other AI enabled lenders and loan servicers of the latest vintages is that the credit markets are currently skewed by the actions of the FOMC. The chart below shows loss given default on bank-owned credit card loans.


Credit spreads are compressed and net-default rates on secured assets such as mortgages and auto loans, for example, are falling and in some cases negative. Credit, at the present time, seems to have no cost, as we saw in 2004-2005. Thus measuring the stressed performance of the nouvelle cuisine production of UPST and other fintech lenders is difficult or impossible.

As and when the great credit correction comes, however, we suspect that UPST and other lenders and servicers that have been created since 2008 will be put to the test and, in some cases, fail rather spectacularly. While UPST states that investors in its loans and asset backed securities (ABS) have no recourse to the lender, this same contractual reality did not prevent investors from demanding early redemption from originate to sell lenders such as Wachovia and Citigroup. The firm does have repurchase exposure on $6.2 billion in loans sold to investors, of note.

Goldman Sachs (NYSE:GS) took UPST out at $20 per share at the end of 2020, but since then the stock has traded as high as $170 but closed on June 16,2021 at $119 or just shy of a $10 billion equity market cap. Is 30x book value a ridiculous valuation for this provider of warm consumer loan leads to banks? Yes, we’d say so, which is probably why the stock has lost 40% of its value in the past several months. Without QE as the lubricant, this model would not exist.

We are told: “Upstart-powered loans originated by bank partners are either retained by the bank partners, purchased by the Company and immediately sold to institutional investors under loan sale agreements, or purchased and held by the Company for a period of time before being sold to third-party investors, or held by the Company.”

As we’ve note above, we’ve heard this story before. Once upon a time, there was a company called Ocwen Financial (NYSE:OCN) that had a visionary CEO who thought algorithms and the precursors of AI were sufficient to handle the servicing of distressed mortgage loans. OCN and many other banks and non-banks were proven wrong in this view. We’ll withhold judgement on whether AI enhanced credit products will result in superior credit results “through the cycle” for UPST, but suffice to say that we are skeptical.

The Institutional Risk Analyst is published by Whalen Global Advisors LLC and is provided for general informational purposes. By making use of The Institutional Risk Analyst web site and content, the recipient thereof acknowledges and agrees to our copyright and the matters set forth below in this disclaimer. Whalen Global Advisors LLC makes no representation or warranty (express or implied) regarding the adequacy, accuracy or completeness of any information in The Institutional Risk Analyst. Information contained herein is obtained from public and private sources deemed reliable. Any analysis or statements contained in The Institutional Risk Analyst are preliminary and are not intended to be complete, and such information is qualified in its entirety. Any opinions or estimates contained in The Institutional Risk Analyst represent the judgment of Whalen Global Advisors LLC at this time, and is subject to change without notice. The Institutional Risk Analyst is not an offer to sell, or a solicitation of an offer to buy, any securities or instruments named or described herein. The Institutional Risk Analyst is not intended to provide, and must not be relied on for, accounting, legal, regulatory, tax, business, financial or related advice or investment recommendations. Whalen Global Advisors LLC is not acting as fiduciary or advisor with respect to the information contained herein. You must consult with your own advisors as to the legal, regulatory, tax, business, financial, investment and other aspects of the subjects addressed in The Institutional Risk Analyst. Interested parties are advised to contact Whalen Global Advisors LLC for more information.

Join the conversation and have a little fun at If you are a new visitor to the Stool, please register and join in! To post your observations and charts, and snide, but good-natured, comments, click here to register. Be sure to respond to the confirmation email which is sent instantly. If not in your inbox, check your spam filter.

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Follow by Email