UPST - Optimizing Income, Risk & Mission

There is a great deal of understandable consternation about the delinquency rate of Upstart ticking up and implications for its future and the effectiveness  of its AI credit engine. 
However delinquency rate in itself is a single variable and cannot be considered in isolation without accounting for **holistic pricing of risk.** 

Pricing of risk is a stochastic optimization problem. It has to be assessed in conjunction with approval rates, interest rates among other factors. 
Upstarts banking partners have the flexibility to tweak the Upstart API for their palatable level of risk. They can optimize to maximize profitability or to minimize risk. Others have alluded to 
the other factors, but it hasn't been put into the context of numbers. I've digged into the KRBA reports and past earnings results to get a clearer understanding. Below, I run through an example.

First as most that follow the company know the approval rates for UPST have been ticking up. So in the third quarter of 2021 for every 1000 applicants, 230 were approved by Upstart's AI engine.
Whereas in the fourth quarter of 2019 for every 1000 applicants, 149 were approved by Upstart's AI engine. **Approval rates have gone up by 50 percent.** 

**Average Loan size & Conversion rates by quarter:**

Quarter      Q3CY19   Q4CY19   Q1CY20   Q2CY20   Q3CY20   Q4CY20   Q1CY21   Q2CY21   Q3CY21
Loan Value   $825M    $1,044M  $1,123M  $164M    $909M    $1,249M  $1,729M  $2,795M  $3,130M
No. of Loans 64K      79K      84K      12K      81K      123K     170K     287K     363K
Avg.Loan     $12.9K   $13.2K   $13.3K   $13.7K   $11.2K   $10.2K   $10.2K   $9.8K    $8.6K
Conv. Rates           14.9     14.1     8.7      15.2     17.4     22       24.4     23

Let us simulate a scenario of 1000 applicants for the two quarters the delinquency rates are considered. A full simulation ought to be for the full duration of the loan. But for simplicity
lets assume the delinquency rate at 5 months and annualize the income for the year. 

**Q3-2019**
Applicants: 1000
Conversion rate 14.9%  (Assuming the same as Q4 2019 as cannot find the same)
Average Loan size: $12,891
Interest rate: 12.5% 
Delinquency rate: a) For Upstart Securitization Trust - 2.33% 

Loan amount = 1000 * Loan Size (12,891) * Conversion Rate (12.5%) =  $1,920,759
Annual Interest earned on the Loan =  $240,094
Loss due to delinquency  = $44,753

Net interest income for institutional buyers  = $195,341
**Normalized Earnings - 10.16%**

—------
Notes - 1) Reference for interest rate and delinquency rate:
[https://www.kbra.com/documents/report/26768/abs-upstart-secu...](https://www.kbra.com/documents/report/26768/abs-upstart-securitization-trust-2019-3-new-issue-report).
2) Interest rate assumed is the total gross excess spread for their institutional buyers as defined in above reports.
—--------

**Q3-2021**
Applicants: 1000
Conversion rate: 23%
Average Loan size: $8,628
Interest rate: 18.69%
Delinquency rate: a) For Upstart Securitization Trusts  - 3.04%
                  b) For Upstart Pass Through Trusts - 4.29% 

Loan amount = 1000 * Loan Size ($8,628) * Conversion Rate (23%)=  $1,984,440
Annual Interest earned on the Loan =  $357,199
Loss due to delinquency  = $85,132

Net interest income for institutional buyers  = $272,066
**Normalized Earnings - 13.71%**

—------
Notes - 1) References for interest rate and delinquency rate:
[https://www.kbra.com/documents/report/61852/upstart-pass-thr...](https://www.kbra.com/documents/report/61852/upstart-pass-through-trust-series-2022-st1-new-issue-report)
[https://www.kbra.com/documents/report/56274/upstart-pass-thr...](https://www.kbra.com/documents/report/56274/upstart-pass-through-trust-series-2021-st9-new-issue-report)
2) Interest rate assumed is the total gross excess spread for their institutional buyers as defined in above reports.
3) I do not exactly understand the difference between Upstart Pass Through Trusts and Securitization Trusts, and why delinquency rates for Securitization trusts are lower than pass through trusts
- but KBRA report reports delinquency rates for them separately and for a **fair apples to apples comparison** the delinquency rates for Securitization Trusts for 2019 (2.33%) and 2021 (3.04% approx)
ought to be compared. But in any case lets go with the worst case number of 4.29% (Pass through trusts) for this analysis.
4) Note the interest rate has increased from 12.5 to 18.69 percent from 2019 to 2021. So a 50 percent increase in both interest and approval rates.
—-------

<b>Comparing Q3 2019 vs. Q3 2021, buyers of Upstart’s security loans are earning 35% more (13.71 in 2021 vs. 10.16 in 2019) on the same amount loaned, even after adjusting the risk of higher
delinquency rate. Additionally Upstarts platform allows them to lend more and automate that process significantly. The total Loan Value itself has quadrupled in 2 years from 2019 to 2021.</b>

<b>In Summary, I think  Upstarts AI engine in 2021 is more effective than the AI engine in 2019. And this is what an effective AI engine is supposed to do, it improves over time!
It has made credit more accessible to the underserved section of society while increasing the profits for its partners.</b>
94 Likes

I missed the previous thread that was closed on Upstart. I think it important to clear up the idea that delinquency rates may be ticking up on newer loan packages.

First of all thanks to jonwayne235 for bringing the loan reports to the board. I looked through them before investing in Upstart months ago.

I don’t believe you can compare the delinquency rates without looking at the underwriting standards set by the banking institutions (Cross River & FinWise) for each loan package. The underwriting standards are different for each loan package and set by the banking institutions.

Summary:The newer loan package has at least 20% of borrowers in a lower credit band and in fact some that don’t have a credit score. The older loan package has a higher minimum FICO score and addition requirement if a FICO score isn’t available. This is all by design by the lending institutions and enacted by the Upstart AI.

More detail on each loan package
2021-ST6 4.29 delinquency rate from the most recent report

https://www.kbra.com/documents/report/51088/upstart-pass-thr…

Page 21 shows a graph that 36-month Upstart loans from Q1 2019 to q1 2021 went from lower 60’s% to 86% in credit grade Upstart’s lowest credit grades (C, D, E and F).

In other words, loan standards were intentionally lowered by the lending institutions on the 36 month loans in the more recent 2021 loan package. Even to the point that a FICO score isn’t required. It’s understandable that delinquency rates might tick higher but that is build into the loan package as the banks are getting a higher weighted APR the lower down the credit spectrum they go. That’s called a win-win for Upstart and the local banks.

Quote from the 2021 report: In Q1 2021, approximately 86% and 68% of the 36 month and 60 month loans, respectively were categorized in Upstart’s lowest credit grades (C, D, E and F)

More detail on each loan package:
2021-ST6 5 newer loan package
Cross river bank FICO requirements: 600 if a credit score is available. No additional requirements if a credit score is not available. Yes that’s right Cross Rive Bank is so comfortable with the Upstart selection process they don’t require a credit score and FinWise’s minimum FICO is 580.
Breakdown of credit spread in this loan package
Grade F credit 3%
Grade e credit 42%
Grade D credit 15%
Grade C credit 14%

2 019-2 5 2.33 delinquency rate from the most recent rating. Older loan bundle that has higher quality credit profile.

https://www.kbra.com/documents/report/21644/abs-upstart-secu…

The lending requirement for these loans in 2019 are more stringent and the APR of the loan package was lower for that reason. The banks will make less money on this loan portfolio than the newer loan bundle if delinquency rates

Wtd Avg Coupon 17.03%
Wtd Avg FICO 691, min 620 or associate college degree and verifiable source of income
Wtd Avg Original Term (mths) 55
Wtd Avg Remaining Term (mths) 51 5

I am still invested in Upstart and believe they can change the lending industry, but I do think it’s going to take longer than I originally thought.

38 Likes

Reposting Oracleoo’s OP formatted for legibility. Thanks to him for the post:

"
There is a great deal of understandable consternation about the delinquency rate of Upstart ticking up and implications for its future and the effectiveness of its AI credit engine.
However delinquency rate in itself is a single variable and cannot be considered in isolation without accounting for holistic pricing of risk.

Pricing of risk is a stochastic optimization problem. It has to be assessed in conjunction with approval rates, interest rates among other factors.
Upstarts banking partners have the flexibility to tweak the Upstart API for their palatable level of risk. They can optimize to maximize profitability or to minimize risk. Others have alluded to
the other factors, but it hasn’t been put into the context of numbers. I’ve digged into the KRBA reports and past earnings results to get a clearer understanding. Below, I run through an example.

First as most that follow the company know the approval rates for UPST have been ticking up. So in the third quarter of 2021 for every 1000 applicants, 230 were approved by Upstart’s AI engine.
Whereas in the fourth quarter of 2019 for every 1000 applicants, 149 were approved by Upstart’s AI engine. Approval rates have gone up by 50 percent.

Average Loan size & Conversion rates by quarter:


Quarter      Q3CY19   Q4CY19   Q1CY20   Q2CY20   Q3CY20   Q4CY20   Q1CY21   Q2CY21   Q3CY21
Loan Value   $825M    $1,044M  $1,123M  $164M    $909M    $1,249M  $1,729M  $2,795M  $3,130M
No. of Loans 64K      79K      84K      12K      81K      123K     170K     287K     363K
Avg.Loan     $12.9K   $13.2K   $13.3K   $13.7K   $11.2K   $10.2K   $10.2K   $9.8K    $8.6K
Conv. Rates           14.9     14.1     8.7      15.2     17.4     22       24.4     23

Let us simulate a scenario of 1000 applicants for the two quarters the delinquency rates are considered. A full simulation ought to be for the full duration of the loan. But for simplicity
lets assume the delinquency rate at 5 months and annualize the income for the year.

Q3-2019
Applicants: 1000
Conversion rate 14.9% (Assuming the same as Q4 2019 as cannot find the same)
Average Loan size: $12,891
Interest rate: 12.5%
Delinquency rate: a) For Upstart Securitization Trust - 2.33%

Loan amount = 1000 * Loan Size (12,891) * Conversion Rate (12.5%) = $1,920,759
Annual Interest earned on the Loan = $240,094
Loss due to delinquency = $44,753

Net interest income for institutional buyers = $195,341
Normalized Earnings - 10.16%

—------
Notes - 1) Reference for interest rate and delinquency rate:
https://www.kbra.com/documents/report/26768/abs-upstart-secu…
2) Interest rate assumed is the total gross excess spread for their institutional buyers as defined in above reports.
—--------

Q3-2021
Applicants: 1000
Conversion rate: 23%
Average Loan size: $8,628
Interest rate: 18.69%
Delinquency rate:
a) For Upstart Securitization Trusts - 3.04%
b) For Upstart Pass Through Trusts - 4.29%

Loan amount = 1000 * Loan Size ($8,628) * Conversion Rate (23%)= $1,984,440
Annual Interest earned on the Loan = $357,199
Loss due to delinquency = $85,132

Net interest income for institutional buyers = $272,066
Normalized Earnings - 13.71%

—------
Notes - 1) References for interest rate and delinquency rate:
https://www.kbra.com/documents/report/61852/upstart-pass-thr…
https://www.kbra.com/documents/report/56274/upstart-pass-thr…
2) Interest rate assumed is the total gross excess spread for their institutional buyers as defined in above reports.
3) I do not exactly understand the difference between Upstart Pass Through Trusts and Securitization Trusts, and why delinquency rates for Securitization trusts are lower than pass through trusts

  • but KBRA report reports delinquency rates for them separately and for a fair apples to apples comparison the delinquency rates for Securitization Trusts for 2019 (2.33%) and 2021 (3.04% approx)
    ought to be compared. But in any case lets go with the worst case number of 4.29% (Pass through trusts) for this analysis.
  1. Note the interest rate has increased from 12.5 to 18.69 percent from 2019 to 2021. So a 50 percent increase in both interest and approval rates.
    —-------

Comparing Q3 2019 vs. Q3 2021, buyers of Upstart’s security loans are earning 35% more (13.71 in 2021 vs. 10.16 in 2019) on the same amount loaned, even after adjusting the risk of higher
delinquency rate. Additionally Upstarts platform allows them to lend more and automate that process significantly. The total Loan Value itself has quadrupled in 2 years from 2019 to 2021.

In Summary, I think Upstarts AI engine in 2021 is more effective than the AI engine in 2019. And this is what an effective AI engine is supposed to do, it improves over time!

It has made credit more accessible to the underserved section of society while increasing the profits for its partners.
"

35 Likes

Reposting my post as the formatting got messed up when using fixed font for tables (sorry had never posted tables). The preview looked perfect so I didn’t realize he issue, sorry for that. I think I have got this right now


There is a great deal of understandable consternation about the delinquency rate of Upstart ticking up and implications for its future and the effectiveness of its AI credit engine. However delinquency rate in itself is a single variable and cannot be considered in isolation without accounting for holistic pricing of risk.

Pricing of risk is a stochastic optimization problem. It has to be assessed in conjunction with approval rates, interest rates among other factors. Upstarts banking partners have the flexibility to tweak the Upstart API for their palatable level of risk. They can optimize to maximize profitability or to minimize risk. Others have alluded to the other factors, but it hasn’t been put into the context of numbers. I’ve dug into the KRBA reports and past earnings results to get a clearer understanding. Below, I run through an example.

First as most that follow the company know the approval rates for UPST have been ticking up. So in the third quarter of 2021 for every 1000 applicants, 230 were approved by Upstart’s AI engine. Whereas in the fourth quarter of 2019 for every 1000 applicants, 149 were approved by Upstart’s AI engine. Approval rates have gone up by 50 percent.


Average Loan size & Conversion rates by quarter:
Quarter      Q3CY19   Q4CY19   Q1CY20   Q2CY20   Q3CY20   Q4CY20   Q1CY21   Q2CY21   Q3CY21
Loan Value   $825M    $1,044M  $1,123M  $164M    $909M    $1,249M  $1,729M  $2,795M  $3,130M
No. of Loans 64K      79K      84K      12K      81K      123K     170K     287K     363K
Avg.Loan     $12.9K   $13.2K   $13.3K   $13.7K   $11.2K   $10.2K   $10.2K   $9.8K    $8.6K
Conv. Rates           14.9     14.1     8.7      15.2     17.4     22       24.4     23

Let us simulate a scenario of 1000 applicants for the two quarters the delinquency rates are considered. A full simulation ought to be for the full duration of the loan. But for simplicity lets assume the delinquency rate at 5 months and annualize the income for the year.

Q3-2019
Applicants: 1000
Conversion rate 14.9% (Assuming the same as Q4 2019 as cannot find the same)
Average Loan size: $12,891
Interest rate: 12.5%
Delinquency rate: a) For Upstart Securitization Trust - 2.33%

Loan amount = 1000 * Loan Size (12,891) * Conversion Rate (12.5%) = $1,920,759
Annual Interest earned on the Loan = $240,094
Loss due to delinquency = $44,753

Net interest income for institutional buyers = $195,341
Normalized Earnings - 10.16%

—------
Notes - 1) Reference for interest rate and delinquency rate:
https://www.kbra.com/documents/report/26768/abs-upstart-secu…
2) Interest rate assumed is the total gross excess spread for their institutional buyers as defined in above reports.
—--------

Q3-2021
Applicants: 1000
Conversion rate: 23%
Average Loan size: $8,628
Interest rate: 18.69%
Delinquency rate: a) For Upstart Securitization Trusts - 3.04%
b) For Upstart Pass Through Trusts - 4.29%

Loan amount = 1000 * Loan Size ($8,628) * Conversion Rate (23%)= $1,984,440
Annual Interest earned on the Loan = $357,199
Loss due to delinquency = $85,132

Net interest income for institutional buyers = $272,066
Normalized Earnings - 13.71%

—------
Notes - 1) References for interest rate and delinquency rate:
https://www.kbra.com/documents/report/61852/upstart-pass-thr…
https://www.kbra.com/documents/report/56274/upstart-pass-thr…
2) Interest rate assumed is the total gross excess spread for their institutional buyers as defined in above reports.
3) I do not exactly understand the difference between Upstart Pass Through Trusts and Securitization Trusts, and why delinquency rates for Securitization trusts are lower than pass through trusts

  • but KBRA report reports delinquency rates for them separately and for a fair apples to apples comparison the delinquency rates for Securitization Trusts for 2019 (2.33%) and 2021 (3.04% approx)
    ought to be compared. But in any case lets go with the worst case number of 4.29% (Pass through trusts) for this analysis.
  1. Note the interest rate has increased from 12.5 to 18.69 percent from 2019 to 2021. So a 50 percent increase in both interest and approval rates.
    —-------

Comparing Q3 2019 vs. Q3 2021, buyers of Upstart’s security loans are earning 35% more (13.71 in 2021 vs. 10.16 in 2019) on the same amount loaned, even after adjusting the risk of higher delinquency rate. Additionally Upstarts platform allows them to lend more and automate that process significantly. The total Loan Value itself has quadrupled in 2 years from 2019 to 2021.

In Summary, I think Upstarts AI engine in 2021 is more effective than the AI engine in 2019. And this is what an effective AI engine is supposed to do, it improves over time! It has made credit more accessible to the underserved section of society while increasing the profits for its partners.

33 Likes

Minor correction:

In the following section the Loan amount line the Conversion Rate should be 14.9% instead of 12.5% and should read like:

Loan amount = 1000 * Loan Size (12,891) * Conversion Rate (14.9%) = $1,920,759

Q3-2019
Applicants: 1000
Conversion rate 14.9% (Assuming the same as Q4 2019 as cannot find the same)
Average Loan size: $12,891
Interest rate: 12.5%
Delinquency rate: a) For Upstart Securitization Trust - 2.33%

Loan amount = 1000 * Loan Size (12,891) * Conversion Rate (12.5%) = $1,920,759
Annual Interest earned on the Loan = $240,094
Loss due to delinquency = $44,753

Net interest income for institutional buyers = $195,341
Normalized Earnings - 10.16%

This is fantastic post. It was the most helpful one on the delinquency questions. Thanks for doing that digging and for clarifying it for those of us who can only offer conjectures.

Can you post again reformatted? I can’t read the text.

You can just format the table within the post and leave the post as is:

< pre > and < / pre > are HTML tags to set off text as preformatted. That means that whatever appears between those two tags will appear in a fixed-width font, formatted exactly the way you typed the text, line breaks and all.

Or if you don’t feel like it, that’s fine…the gist is there.

2 Likes

Oracleoo,
I greatly appreciate this post. With deepest respect to Jon Wayne, who has brought us the benefit of some incredible analysis on Upstart, raised delinquency rates as a red flag. And GauchoChris interpreted it as broken investment thesis, another board member for whom I have a great deal of respect. I could only see it as a pink flag and a phenomena I failed to understand.

The reason I did not perceive it to be a dire warning or broken thesis is because it just didn’t really make sense. What I mean by that is that it only made sense if you believed that the Upstart AI engine had more or less suddenly gone into failure mode. Either that or Upstart management was just lying to us investors. Us and their banking partners who somehow hadn’t noticed that things were headed south. I just couldn’t buy into the notion of deceptive management. And while I am most definitely not an AI expert (AI was just beginning to be a thing when I retired), this is simply contrary to everything I understood about how AI works. Especially the part about increasing accuracy over time as the training data becomes further enriched with real world experience.

That left only one alternative. The increased delinquencies were by design. I didn’t post anything about it because I was at a total loss to explain why Upstart in conjunction with their partner banks would elect to do this. Your post has made that explanation crystal clear (at least it is for me). In a nutshell, tolerating a higher rate of delinquencies and ensuing defaults is more profitable.

Banks like that.

22 Likes

Hi Oracleoo,

This is a fantastic post. I wish I could recommend it several times. Thanks for analyzing and sharing your thoughts.
I agree completely with what you laid out. As also detailed by other posters earlier, it is very clear that delinquency rates will be higher in the 2021 trusts than 2019 trusts by design - they are incorporating more borrowers with much lower FICO scores or other criteria that make them less trustworthy, at least according to traditional modeling. It’s also clear that riskier loans will have higher interest rates and more potential profit to the lender, despite higher delinquency rates.
What is not clear to me, however, is whether the increased expected earnings as adjusted for the higher delinquency rate is any better than if similar loans were underwritten by traditional modeling/FICO scores.

I believe what also must be considered, is whether Upstart’s 2021 trusts are still outperforming what would be expected out of them by traditional modeling. Because if Upstart’s delinquency or default rates are not much different than what FICO predicts, then their AI does not truly have any advantage to the status quo.
And just as GauchoRico said before, “the competitive advantage of better risk assessment and therefore better underwriting and therefore lower rates for borrowers AND lower delinquencies for lenders is Upstart’s reason for being.
Why should more banks choose to partner with UPST, if UPST cannot prove their underwriting will lower defaults significantly than whatever models they normally use?

Now, let me be crystal clear. If UPST is the same or no better than FICO and traditional modeling to predict defaults, it doesn’t mean that UPST will go out of business.
UPST can still provide value to bank partners by their superior AI marketing models and digital platform that many paper-based banks/credit unions completely lack. UPST has a good brand to reach customers out of bounds for many banks - they can still be a nice additional referral network for its partners. UPST is still a good conduit to flow capital from institutional investors through loan securitizations.
I mean, for example, lendingclub or lendingpoint or Upgrade or whatever are still existing and growing today, right? And we know their past loan securitizations don’t beat out UPST’s pre-2021 loans, but they haven’t gone bankrupt.
BUT if UPST’s underwriting AI models can’t outperform traditional ways of predicting risk, I believe it becomes a huge risk and barrier to sustaining long-term HYPERgrowth.
Eventually, defaults/delinquencies will catch up to UPST if they try lending too loosely and no better than traditional FICO methods - basically, in the long run, their total loan volume will be constrained by the failure to find true-hidden prime borrowers - the same exact way it has constrained Lendingclub (they grew very quick in the beginning, then loan volume growth dropped like a rock when defaults piled up).

I’d like to share what I found after carving some time out yesterday to dig deeper into the 2021 pass through trusts. I shared the below findings with some great investors yesterday and decided I’ll post it here for anyone else to see too - I’m looking for what everyone else thinks and any other feedback:

I decided to compare KBRA’s initial forecasted cumulative net loss projections versus the actual cumulative net loss. I think this is a much more ‘apples to apples’ comparison than simply using delinquency rates across different year vintages, which is what I first posted on the board. This way, I think we strip out confounding factors of different interest rates/loan sizes/avg FICO scores etc.
Now we are using a standardized measurement of expected loan performance to itself - presumably, KBRA’s loss projections are created by traditional FICO modeling that take into account all the usual variables [SEE FOOTNOTE 1 AT BOTTOM OF POST].
So if UPST’s AI is truly better, then UPST’s 2021 trusts should outperform the initial KBRA loss projection in all vintages (like UPST was indisputably proven to be doing in all trusts before 2021)

Here is what I found (screenshots below):
https://ibb.co/mXz7N7M 2019-1 trust, marked up by me
https://ibb.co/X83xqKJ 2019-1 trust, unmarked
https://ibb.co/m07J86t 2021-ST3 trust, marked up by me
https://ibb.co/G3rxt3X 2021-ST3 trust, unmarked

For UPST pass through trust 2021-ST3, which the date of closing was April 2021, I took a snapshot of KBRA’s projected base case cumulative net loss graph. I then looked at January 2022’s latest KBRA report which indicates that at 8 months of seasoning, 2021-ST3 was experiencing a cumulative net loss rate of 2.27%. I plotted this as precisely as I could by using the pixel counts in microsoft paint, onto the snapshot of the base case CNL graph.
What I found: at 2.27% for 8 months of seasoning, UPST 2021-ST3 is exactly in-line with its initial loss projections [in the screenshot, 2.27% at 8 months intersects exactly with the initial projection loss curve]. I do not view this as a good thing.
As we know from every single trust prior to 2021, UPST has outperformed each initial base case loss projection by huge margins (remember, they were all ~40% better).

To verify this I also looked at UPST 2019-1 trust. [Why look at a prior trust? Well, it’s just to make sure that CNL rates aren’t typically “just in-line” to KBRA initial projections in the first 12 months or something, and then suddenly drop dramatically later on by month 20 or whatever, that results in the final CNL as 40% outperforming the initial KBRA projection. This removes that possible confounding factor at least.]
I chose trust 2019-1 because its closing date was February 2019 and so at 7 or 8 months out its losses will have zero COVID or stimulus impact whatsoever to confound the picture.
So with the same methods above, after plotting the CNL at 7 months of seasoning, we can see very substantial outperformance, against the initial base case loss projections.
UPST 2019-1 had loss of 1.97% while base case was measured at 4.16% (using the same MS paint pixel measurement method). That means UPST 2019-1 outperformed what was expected, by 52.6%. That’s a far-cry from 0% outperformance for UPST 2021-ST3!

What I’m seeing from this sampling, is that 2021 trusts are probably not really doing any better than traditional modeling/FICO scores would suggest. This is quite problematic and seals the deal to me that this is not just uncertainty, but a truly confirmed red flag for UPST’s recent AI performance. Like GauchoRico has said, this can break the investing thesis.

It is possible that the reason they had such explosive 61% QoQ and 38% QoQ growth in the first half of 2021 was by lending way too loosely - maybe their AI was no longer finding “true hidden primes”, but just tossing loans at borrowers “willy-nilly”. Will they get better in the future and course correct? Maybe.

FOOTNOTE 1:
You can read how KBRA comes up with its initial base case loss projections: https://documents.krollbondratings.com/report/184/abs-genera…

They will factor in:
-Obligor credit scores - internal credit scores and/or externally produced credit scores, internal or external ratings
-Product or loan type
-Contract original and remaining term / amortization schedule
-Interest rates / yield
-Leverage ratios, as applicable, such as debt-to-income (DTI), payment-to-income (PTI), debt service coverage,
down payment, and loan-to-value (LTV) ratios
-Geographic distribution
-Specific collateral and loan characteristics relative to that asset type

And, they also try to factor in macroeconomics:
…For a given transaction, there may be certain adjustments made to account for seasoning, changes in the economic environment and jurisdictional considerations or other factors.

FOOTNOTE 2:
One possible explanation for the lack of outperformance in UPST-2021-ST3 is that KBRA already took into account UPST’s supposedly superior AI underwriting. Personally, I don’t think this is the case, because all of UPST’s trusts before 2021 have outperformed by huge margins - and it just suddenly is not doing so in 2021. If KBRA is actually factoring in UPST’s superior AI over time, then why wouldn’t the spread between the initial base case and actual outperforming CNL results not narrow from 2017 to 2020? (the outperformance margins have been consistently huge, see screenshot: https://ibb.co/g9nXYfJ)

FOOTNOTE 3:
I feel like this doesn’t need to be said, but this caveat is just as important. I’m a regular person with no special skills or information. All of the above that I say could be very wrong or incorrect. The best way we can know is with time to see how UPST plays out. Maybe in Q4 they will explain these things satisfactorily. Or future KBRA reports show they start to outperform again. Or something else can be understood here that I’m not smart enough to think of. Either way, it’s too uncertain and I am definitely staying out and watching from the sidelines at this point.

29 Likes

JonWayne,

First, excellent second look and we are all more knowledgeable because of your research. Your additions have already led me to be a better investor.

In terms of KBRA adjusting the underwriting, to me that is obvious. If you look at the graphs linked, 2019-1 trust had an expected loss at month 60 is 21.1%. In 2021-ST3 it is 19.3%.

This is despite approximately 52% of loans in this package being D, E, or F loans compared to 2019-1 which only comprised of 76% A, B, C loans, 13.6% D and 10% O/C (unclear to me what this is).

2019-1: https://documents.kbra.com/report/16145/abs-upstart-securiti…

2021-3: https://documents.kbra.com/report/51522/upstart-securitizati…

I encourage everyone view the historical performance section of 2021-3 (pg. 20/21) and it will show the historical performance of each loan grade and term. It shows in Q1 2019 (36 mo. loans) there were approximately 35% AA, A and B whereas in Q4 2020 was 13.74%. In 2022-1, that ticked up slightly to 14.45%.

Now lets look at 2022-1 for this set, the base case is 16.75%. This batch of loans is comprised of 36, 60 and 84 month loans. 57% are D, E, or F.

So to recap:
2019-1: 76% ABC loans, expected loss 21.1%
2021-3: 48% ABC loans, expected loss 19.3%
2022-1: 43% ABC loans, expected loss 16.75%

One more data point:
2021-5: 50% ABC loans (almost all 36/60 mo), expected loss 15.15%

This leads me to believe Upstart is funding “riskier” loans, but yet the rating agency is projecting less loss over time. To me, this is a reflection of their AI working and the expected loss would have to factor in the known delinquencies earlier in the report.

The rating agency is getting more comfortable with giving Upstart better ratings. I parsed through the exact data, but your image also shows this trend of the rating agency, that over time from 2017 to 2020 ST-3 the general trend for the base case is down (there will be fluctuations based on the make up of the package). Since none of the 2021 loans have been reviewed there is no information as to the current loss projection. If these go above the base case, then the thesis may be broken. Similar to expecting a huge beat against guidance (which they indicated they are trying to be accurate), as they have more longevity and data, I would expect base case and actual loss to get closer and closer. I assume Upstart would be able to charge a higher premium for securitization trusts that have a higher interest rate and lower expected loss, but I am not 100% sure in how these are marketed/sold.

I, too, am just an ordinary person trying to make the best investment decision I can. I still feel the thesis is in tact and the most important metrics outside of revenue growth, will be new CU/bank partners and rooftops added, and for me Auto loans having meaningful contribution.

Thank you to everyone on the board. I hope this post is useful.

JE

46 Likes

Hi JonWayne,

Thanks for your great research over the months, we have all benefited from it.

FoolishJe has already made some excellent points along similar lines that I was about to make. I will add one more to it on your assumption made in Footnote 2:

One possible explanation for the lack of outperformance in UPST-2021-ST3 is that KBRA already took into account UPST’s supposedly superior AI underwriting. Personally, I don’t think this is the case, because all of UPST’s trusts before 2021 have outperformed by huge margins - and it just suddenly is not doing so in 2021.

Actually based on the report below KBRA do take into account UPST’s past historical performance to develop base case loss expectation.

https://www.kbra.com/documents/report/61852/upstart-pass-thr…

On Page 24:
KBRA Loss Expectation KBRA has analyzed Upstart’s historical static pool cumulative gross loss data since the Company began originating loans, segmented by term and Upstart grade in order to develop a base case loss expectation per KBRA’s Consumer Loan ABS Global Rating Methodology…This proxy data, in addition to Upstart’s management loss analysis was used to help derive and validate the base case loss assumptions for each of the loan grades and original loan terms

18 Likes

Hi JE,

Great thought. I considered the same thing when I noticed the 21.1% initial expected CNL for 2019-1 was higher 19.3% for UPST-ST3. But yesterday I drew a much different conclusion when I took a granular look at each report.

First, I disagree with your calculation of 2019-1 having 76% ABC loans, it’s not what’s in the report. In 2019-1 new issue report, you can add up the tiers (on page 25) to come up with 48.98% are ABC loans. Not 76%.
In 2022-ST3 new issue report, you can add up the tiers (on page 11) to come up with 38% are ABC loans. Not 48%.
So the spread between the two in terms of creditworthy tiers is much smaller than you had calculated.
If I am somehow adding up the numbers incorrectly please look at the pages I cited and let me know.

Second, the mix of loan terms are way more favorable for 2022-ST3 and NOT for 2019-1. We can see 2019-1 had 83.81% sixty-month term loans, 15% thirty-six month term loans and 1.09% eighty-four month term loans.
2022-ST3 has 65% sixty-month term loans and 35% thirty-six month term loans.

Longer term loans tend to be much riskier as loan balances are much higher (if a borrower defaults, a much greater principal amount goes unpaid. You are placing larger but fewer eggs in your basket, so to speak), and much greater amount of time to produce defaults (a 5 year loan is also much more likely to encounter a recession/economic cycle than a 3 year loan).
I mean this is why Afterpay’s default rate is way lower than its competitor Klarna. Afterpay’s finance product is pay-in-four installments over 8 weeks. Klarna offers some products for up to 36 months and thus experiences higher defaults.
Anyway, you can see that 2019-1 has a lot higher proportion of longer term (and therefore riskier) loans than 2022-ST3. So it makes perfect sense to me why 2019-1 would produce a higher initial base case CNL higher than 2022-ST3, all else being equal. Except not all else is equal! Too many factors are different and this is apple-to oranges.

Third, now you have to take into account the differences in overcollaterization and gross excess spreads. There’s better spread for 2022-ST3 but better overcollaterization for 2019-1. How do we reconcile this? I don’t know, I’m not a finance expert.

As I said earlier:
FOOTNOTE 2:
One possible explanation for the lack of outperformance in UPST-2021-ST3 is that KBRA already took into account UPST’s supposedly superior AI underwriting. Personally, I don’t think this is the case, because all of UPST’s trusts before 2021 have outperformed by huge margins - and it just suddenly is not doing so in 2021. If KBRA is actually factoring in UPST’s superior AI over time, then why wouldn’t the spread between the initial base case and actual outperforming CNL results not narrow from 2017 to 2020? (the outperformance margins have been consistently huge, see screenshot: https://ibb.co/g9nXYfJ)

I think my concern still holds completely true! I believe when you try to explain away the lack of outperformance in 2021-ST3 compared to vintages from 2020 and earlier, it is becoming way too much of an apples-to-oranges comparison when you try to compare the CNL from one loan vintage to another. There are too many factors at play, too many factors favoring 2019-1 and other factors favoring 2022-ST3 as I detailed above.

Finally, I disagree with your statement “but your image also shows this trend of the rating agency, that over time from 2017 to 2020 ST-3 the general trend for the base case is down”
All I am seeing is that the margin of outperformance is as follows (from the screenshot table above):
45%, 41%, 36%, 39%, 43%, 44%, 38%, 71%, 66%, 48%, 56%, 54%…and then suddenly ZERO percent outperformance in 2021-ST3.
That just doesn’t look good to me. The outperformance spread wasn’t narrowing at all over time! (Please calculate and let me know if I got the number wrong. I am taking “current period CNL” divided by “expected base case CNL for current period” to find out the margin of actual outperformance.)

If I can’t rely on apple-to oranges comparison of comparing CNL to CNL on different loan vintages, then I need to rely on initial loan performance vs actual loan performance that strips out all the confounding factors mentioned above. And all I’m still seeing here is a red flag.

Also, not related to the details of the above post, but I will paste the newswire that came out Friday afternoon from Wedbush on my brokerage. It looks like the wedbush analyst has access to Bloomberg terminal data that we don’t, unfortunately! If we did we might be able to compare things with greater ‘real time’ information.

Upstart Holdings’ (UPST) increasing challenges with respect to delinquencies could impact its credit-adjusted return, Wedbush said in a note to clients Friday.
Based on an average of data from Bloomberg for 2021 vintage securitizations, delinquency rates rose to 2.57% in January from 2.44% in December, according to the investment firm.
“The average delinquency rate through [eight] months for 2021 vintages is 2.57%,” said Wedbush, which noted an overall average delinquency rate of 1.45% through eight months for all vintage securitizations.
The delinquencies and related loss rates could affect investment return, the firm said, noting the consumer lending company’s target credit-adjusted return of 6% to 7% for its credit investors. However, it expects revenue trends for Q4 and early 2022 to still be strong before this takes effect.
“Furthermore, the true test will be if Upstart’s underwriting model withstands an economic downturn with better-than-expected losses, and if achieved, this could set the stage for the company to make significant strides in gaining market share,” Wedbush said.
Wedbush lowered its price target for Upstart to $110 from $160 while reiterating its neutral rating.

7 Likes

JE, I just noticed why your numbers were not correct for 2021-ST3 regarding ABC grade composition. You were looking at 2021-3, which is a totally different trust than the previously referenced 2021-ST3.

Also I had typos in my reply to you. Wherever I kept saying 2022-ST3, I was meaning 2021-ST3. My mistake.

2 Likes