Monday, September 6, 2021

Upstart (UPST) and its AI lending, is it working? a real world impact

DISCLAIMER: I am not a financial person. Analysis below is based on my understanding of the complex financial jargon and terms. If you find any discrepancies and wish to connect, find me on twitter (handle in figure). All data is collected from publicly available sources. I own shares of Upstart Holdings Inc.

INTRODUCTION: Upstart is an AI lending platform that partners with banks and credit unions to provide consumer loans. I wrote an article few days ago outlining the financial ecosystem that Upstart has created. It uses a proprietary  artificial intelligence algorithm to predict creditworthiness. I analyzed data from Kroll Bond Rating Agency (KBRA) regarding Upstart to see if it was indeed doing what it was supposed to do. Following is the charting and analysis of that data.


Consumer loan Asset backed securities (ABS): financial securities backed by income-generating assets.

Weighted average interest rate: The interest rate is the amount a lender charges a borrower. Weighted average is for the loans in the collateral pool for that cut-off period.

Weighted average FICO score: Weighted average FICO (Fair Isaac Corporation) score for borrowers for the loans in the collateral pool. This is a measure of creditworthiness. Higher FICO score means borrower will have a lower interest rate on loan and vice-versa.

Weighted average original term: Weighted average of loan term of the loans in the collateral pool. Longer term loans are for those who want to borrow large amounts or want to have a lower monthly payment. 

Weighted average seasoning: Seasoned issues tend to be associated with less risk and are more favorable. An entity dealing with loans would like duration for seasoning to shorten making the loans more favorable.



A Temporal analysis of Upstart lending parameters (click on image to zoom)


Every successive assessment by KBRA shows that consumer loan ABS transactions have increased from 65 million to 485 million, which is a 646% increase from 2020-2 to 2021-3 time period (Panel A, top left). This indicates Upstart has been exceptionally successful in securing increasing number of loan transactions.

 Borrowers in Upstart have an increase in weighted average interest rate panel B (top middle) Most lenders look at borrower risk—how likely you are to pay back the loan. Upstart is facilitating lending to borrowers at higher interest rates, indicating these are borrowers who would not otherwise qualify for a loan. Potential borrowers with higher credit scores tend to get more favorable (lower) interest rates, so the fact that this is trended higher but stabilized at around 18%. Based on data from Experian average interest rate on a personal loan is 9.41%, according to Experian data from Q2 2019. Depending on the lender and the borrower's credit score and financial history, personal loan interest rates can range from 6% to 36%. Thus the rising interest rates offered are a glimpse into Upstarts algorithms approving additional borrowers.

Upstart had  a reduction in the FICO scores of the borrowers over time (Panel C, top right) which in my opinion, speaks to the effectiveness of the AI algorithms that Upstart uses to offer loans. I expect this to keep dropping until it reaches a lower limit, as the AI keeps getting better and better at assessing creditworthiness. For those banks that would use FICO, these borrowers with lower scores would either not qualify at all or would qualify at very high rates. However, Upstart would be able to tap into a huge pool of potential borrowers who have low FICO scores (thus penalized by a non-upstart lending entity) since its AI would identify credit worthy borrowers in this pool. Also note that a score 670 and up includes 'Good', 'Very Good' and 'Exceptional' categories (captures 66% of American Market), while a score of 580 and up includes 'Fair' along with other groups and captures 84% market. That Upstart is able to identify worthy borrowers for loans with 'Fair' score is a credit to its AI based determination of creditworthiness. Whether or not it will dip to 580 or below is not know, but what is clear is that its algorithms capture a larger market share than conventional lending.

The weighted average original term (Panel D, lower left), which is the loan term dropped in 2021-1 assessment likely a reflection of COVID-19 effect on borrowing. However, this has quickly bounced back and gone up to 55 months from a starting of 53 months. Longer term loans are for those who want to have a lower monthly payment. This anticipate this will also rise until a limit and would be a reflection of those borrowers who qualified, could pay but prefer lower monthly payments. The weighted average seasoning of the collateral pool has also dropped to 4 months from a starting value of 8 months. Thus the time to 'season' has reduced. An entity dealing with loans would like duration for seasoning to shorten making the loans more favorable. 

Thus based on the KBRA data, Upstart is making more loans to an increasingly FICO score marginalized population. Which means, its AI algorithm is working as planned and is actually improving every successive iteration. This in my opinion is a better solution to solve a huge problem for human beings making UPST a disruptor stock.