Superior Risk-adjusted Pricing: UPST stated differentiation.

Recall from our previous note that one of the key elements of UPST’s economic model is that its superior proprietary artificial intelligence credit scoring model allows the company to generate higher approval rates with lower interest rates and better terms and, yet, at the same time provide better risk adjusted returns to its network of lending partners who make the loans.

UPST’s proprietary credit scoring model, like all credit scoring models, grades borrowing applicants from best to worst according to expected default risk.  Based on these default expectations, UPST determines the interest rate that would sufficiently compensate for the risk of funding this loan. That is to say, in the absence of offers from competitors[1], UPST could theoretically be profitably underwriting loans across the full spectrum of its grading scale, from its highest to its lowest grade[2].

But in reality, UPST does face competition.  UPST’s profitable opportunity set, is therefore determined[3] by the extent of the differentiation between UPST’s view of expected default risk and its competitors view of expected default risk. This is UPST’s self-declared Leitmotif.

At a given level of a competitor’s FICO credit score, UPST claims that its credit grading methodology identifies a subset of applicants with lower default risk than what competitors glean from the FICO score alone. An example being an applicant that competitors—relying solely on FICO scores—rank at the low end of their scale (say a FICO score below 640) but UPST’s methodology grades them in the middle of their scale (say C) or better. Probabilistically, UPST’s profitable opportunity set increases as the FICO credit score worsens or as UPST’s credit grade rises.

UPST has been reinforcing this credit differentiation capability to equity investors with a slide in its earnings release presentation ever since 1Q 2022. Figure 1 is the slide from Q2 2023’s earnings results. Management calls out the fact that the average annualized default rate for their worst grade (E- at 17.6%) is 11x the rate of its best (A+ at 1.6%), whereas for traditional FICO score the worst grade (639 or below at 16.4%) is only 4x the rate of best (700 or above at 4.5%). This greater range in outcomes on UPST grading scale versus FICO, they claim, is evidence of their model’s better differentiation compared to a FICO only credit score.

Figure 1

As previously noted, this claimed superior differentiation is the essential basis for a profitable opportunity subset vis-à-vis competition, which is reinforced by inference in this matrix. For example, with its differentiation, UPST borrowers with 700 or above FICO scores and a B grade or better on their scoring methodology achieved default rates lower than the FICO band’s 4.5% average, while those with a C grade or worse produced a higher default rate. The implication being that UPST could have offered 700 or above FICO borrowers graded B or better, lower rates than competitors and still would have provided adequate or better than expected risk adjusted returns for funding partners or for themselves. The ability to offer lower rates would have increased the likelihood of winning more business.

For the 700 or above FICO and C-graded or worse borrowers, UPST would have needed to extend credit at rates higher than competitors which reduced their chances to win that business or if they matched competitors’ rates, it resulted in inadequate risk-adjusted returns for their funding partners or for themselves which is not sustainable beyond the short term. Similarly, such profitable opportunity subsets existed at each FICO score level, but we note (implied from Figure 1) that the focal demarcation line is “C” grade or better for borrowers with FICOs of 660-699, and “D” grade or better for all FICO scores below 660.

Logically, the scope for UPST’s potential per loan contribution profitability is a function of the difference between the “generic” average default rate for a given FICO score range, which presumably competitors use to assign an interest rate loans to the entire band, and the specific default rate of the UPST grade within that FICO score band. Which means the higher the UPST grade within a given FICO score band the higher the per unit potential contribution profitability. The actual per loan contribution profit or spread would depend on how much of the default rate differential is given to the borrower (in the form of lower rates) versus how much is given to the funding partner (in the form of lower loan sale price/facilitation fees for those loans sold). These facts taken together imply that as you move down and to the left on the matrix, the more “desirable” the borrower would be to UPST.

To simultaneously maximize origination growth and total profit, you would expect UPST to target as many “A+” graded borrowers within the lowest FICO band as possible, then as many “B” graded within that FICOs band as possible, and so on. Borrowers are increasingly less desirable to UPST as they move up and to the right in Figure 1 because UPST’s “unsophisticated/undifferentiated” competitors will unwittingly offer borrowers uneconomically lower rates. Equity investors, if they understand fully the implications of this matrix would anticipate UPST underwriting little if any “E” graded loans regardless of FICO score. To do so would completely negate the stated value proposition of UPST’s “AI-driven” competitive advantage.”

Does the matrix really support UPST’s differentiation claim?

To us grizzled financials analysts, a more careful examination of the matrix is less compelling evidence of differentiation and is telling us a different story. To start, as noted in a footnote in UPST’s earnings presentation, the annualized default rates in the matrix are for “…all [UPST] originations made 2018-Q1 to 2023-Q1 vintages.”  Meaning that all the figures in the matrix are the historical annualized default rates experienced on the loans UPST originated over a 5-year period. For example, the 13% default rate shown for E- UPST graded loans in the 700 or above FICO band is not using an estimate of default rates for that FICO score and UPST grade combination but an actual average experienced on loan applicants UPST originated with those characteristics. Likewise, the 4.5% average for the 700 or above FICO band is the historical average UPST experienced and not what the theoretical FICO only lender experienced.

So, this matrix is actually just a report card of realized default rates on what UPST was able to originate over 5 years. We don’t know what the underlying UPST grading distribution of potential borrowers is, nor does it even disclose what the actual origination distribution outcome was by FICO/UPST grade bucket. This data is the average of 5 years of vintages. You’ve got default rates on old vintages that benefitted from low rates/COVID policies and are paid off by now mixed with recent vintages that have yet to season into their default windows. Furthermore, given their AI/machine learning approach UPST have been constantly tinkering/improving with their own grading system. So, an “A+” grade in 1Q 2018 is not the same as an “A+” grade in 1Q 2019 let alone in 1Q 2023. To be of any support or evidence of their claimed differentiated credit scoring the realized default rates must be disaggregated into quarterly by sub-component vintage curves and compared to appropriate industry average FICO curves. If UPST actually has a differentiated superior credit methodology they should have and should be, not only willing, but eager to disclose quarterly vintage curves and comparable industry average FICO curves.

All we can say is that an examination of Figure 1 clearly indicates UPST did not originate equally across the sub-components within a given FICO band.  For example, for the 700 or above FICO band the average is 4.5%, which is a little above the mid-point between the default rates for the B and C grades in that band. Given there are 5 UPST grades sub-components within each FICO band, if UPST originated equally across the 5 sub-components the FICO band, the average default rate would have been 6%: the result of the C grade sub-component. This means that the C, D, E- UPST sub-components collectively within the 700 or above FICO band combined are underweight while the A and B sub-components collectively are overweight.  This makes sense. If UPST grading methodology outlines a sub-component that they believe can profitably undercut competitors rate offerings in the lowest risk FICO band, they should have originations disproportionally in their best credit grades. The puzzle for us is why would they do any originations in that FICO band at their weakest rating grade (E-). A similar analysis illuminates that most of UPST A+ grade originations were concentrated in the 680 or better FICO scores and less so below that. Again, this seems to suggest that while the per unit profitability may be much higher at lower FICO scores, there is a dearth of actual volume of A+ borrowers available in the subprime FICO bands (which is not surprising). Hence, the relative overweight in the higher FICO bands. UPST’s apparent distribution of FICO scores within the E- grade is concentrated in the lower subprime FICO bands.

Collectively, admittedly, from the sparse information in the matrix, it appears that UPST’s actual origination mix by FICO/UPST grade is significantly less a function of the differentiated view of default expectations than it is influenced by UPST’s need to ramp volumes to buttress equity valuations. Said differently, UPST is not getting its loan volumes from where its credit scoring is most differentiated from FICO only lenders, but rather incongruently from where UPST’s grading is least differentiated. Why? because to paraphrase bank robber Willie Sutton, because that’s where the volume is.

To get a better understanding of true credit underwriting at UPST we gathered data from the only source available: the securitization trusts of their loans. Collectively the loans collateralizing the securitizations made post-1Q 2018 total $6.6 billion or 20% of the cumulative $32.7 billion originations during the 5-year period. We can’t claim it definitively represents all of their originations during the period, but it is hard to imagine it could be materially different.

Chart 1

As cynical as we can be sometimes, even we are surprised to see how much of UPST’s securitized loans originated were E- graded loans. Of the $6.6 billion securitized from August 2018 to July 2023, nearly 34% were E- graded loans whereas only 11% were A+ graded loans.  How can a company that claims superior credit underwriting ability profitably grant loans to borrowers that they themselves grade as the highest risk on their own grading scale?

Table 1

Chart 2

To get a better insight, an examination of a specific securitization trust is in order. UPST 2022-1, comprising a $503 million portfolio originated in the 4^th Quarter of 2021, the peak of opportunity (i.e., rates had yet to rise with targeted borrowers still awash with liquidity and buoyed by student loan repayment moratorium).

Table 2

Table 2 provides further support for our view that while some of UPST’s lending is done on the basis of their grading indicating lower relative risk than FICO scoring, for the bulk of its originations, that is not true. In fact, the bulk of UPST’s originations are in parts of its matrix that offer the worst opportunity for better risk-adjusted returns.

We believe that these trends indicate that UPST is taking loan volume wherever it can get it, that it’s doing so despite its claims related to the accuracy of its models, and that this practice is unsustainable because it can’t underwrite E- loans in this rate environment without brushing up against usury rate limits, and it can’t be done on a cash-flow positive basis because of the credit enhancement requirements that will need to accompany these loans. Because of this negative confluence, UPST has little to no ability to reach sustainably profitable loan volume levels.  

[1] And in the absence of legal usury interest rate limits.

[2] There is also the non-trivial matter of adverse selection, where the required offered interest rate is so high that the only borrower applicants that accept the rate are those that will default.

[3] UPST’s profitable opportunity set could be determined by relative funding costs, underwriting speed, branding etc. But UPST has a relative funding disadvantage vis-à-vis traditional incumbents, whereas the meaningfulness and sustainability of underwriting speed and branding is negligible at best.