[112012123]

Interesting. Overall, I agree with everyone else - Affirm looks like a healthy company.

Major takeaways:

1.5% write-off rate for their jan 2020 vintage is very healthy - comparable to the long-term trend for unsecured superprime consumer debt. Given the (I suspect) lower average creditworthiness of Affirm customers, this is a great number. I'd be curious to see their long-term trend for same-age vintages, however. In consumer credit it's well known that all the stimulus support in 2020 has significantly depressed defaults. It would be interesting to see if this is a fluke, or if this is actually what their charge-off rate actually looks like in a normal environment/part of a bigger trend.

I'm a little skeptical of their claim to use ML & build a data moat for significantly better underwriting decisions. Consumer credit laws in the US so severely restrict what you can use for credit scoring purposes that better underwriting through data is basically a lost cause, absent some specific customer segment that has special credit situations.

Finally, as others have noted, 30% of revenue just from Peloton is an enormous number.

[tristor]

> absent some specific customer segment that has special credit situations

It seems from the S-1 that Affirm has grown it's market quite a lot, but my first and primary exposure to them is in the car scene. Affirm entered this customer segment and dominated pretty quickly as the creditor integration of choice for long time-scale projects with up front costs like custom engine builds. I imagine to some degree they are able to make decisions partly based on the nature of what is being purchased.

Their S-1 indicates this is shifting, but I have historically seen Affirm in places where a more typical online consumer credit organization like PayPal Credit / BillMeLater / etc doesn't play. Affirm seemed to be focused on larger sized purchases which would otherwise be paid on an installment plan, but filling that gap. Exercise equipment like Peloton is a great example, just as engine builds is a similar type of transaction.

[raiyu]

When you look at 30% of revenue from Peloton, and that their interest rate revenue is lower than merchant fees, it's basically companies paying to finance larger purchases for their customers because they get to book the full revenue on their books immediately, pay down the interest through merchant fees themselves, get higher numbers on their books, increase their market cap, and for consumers it's a benefit because why not finance it over a period of 39 months than be out of pocket immediately or finance it through a credit card with high interest rates.

So really it's a great way to take advantage of the credit markets and public company comps being extremely high while benefitting the customer.

[momokoko]

> Consumer credit laws in the US so severely restrict what you can use for credit scoring purposes that better underwriting through data is basically a lost cause, absent some specific customer segment that has special credit situations.

Can I inquire what your background is or where you found that information? Many companies supplement credit scores with additional data to make these types of decisions.

[Guest42]

I’ve worked in credit risk modeling and it is rather strict the predictors that can be used and well documented. Data comes in from a variety of sources and it is favorable to be skilled in established models than to try something obscure that isn’t intuitive. The models have to work across different sets of time and the varying business processes that may have been in place. Fraud modeling is more flexible, but seemed to have fairly similar results although more trendy things like random forests and neural nets would show up.

[tadkar]

That is definitely changing in credit risk and underwriting as well. There are several companies like [1] applying deep neural nets to the credit risk problem. This on top of a lot of in house work in the big banks to “supplement” what is available on the open market

[1] https://zest.ai/

[lcrmorin]

There is a big difference between dumping a dataset in the latest hot ML model and building something that offer some actual explainability, deal with intrinsic biases, have some stability over time and recalibration and that will go trough an unprecedented crisis for which you don't have any data to learn from. That mean the model usually has to go trough a lot of internal commities and different external agencies. I highly suspect that an external proprietary solution won't go very far in the credit rating field.

[112012123]

This is definitely a super interesting company/approach - thanks for the link! I'm definitely curious as to whether they're actually using things like neural nets (or any other more-sophisticated technical techniques). The traditional problem is that those models aren't explainable, and potentially have hidden biases in them, so I'd be really curious what their approach is.

[Guest42]

Do you work for them? What is your background in academia/business?