What Happens When Machine Learning Finance Models Fail

These are some strange times. Look, we have $16 trillion of negative yielding bonds, that’s T, for trillion. I’m asked by non-financial people why anyone would want to buy negative yields (you pay to hold them, btw) and I reply, it’s not about income, it’s about trading that rates will fall further.

Which got me thinking: If we are in some liquidity trap world and negative interest rate environment, what does that do to all these invoice financial models being built using the latest and greatest in artificial intelligence and machine learning?

For some insights, I look back at 2007-08, a period when things were getting weird, and that was the mortgage backed security (MBS) crisis, perhaps the biggest market disaster in terms of leverage and securitization tied into securitizing non-conforming mortgages and selling them to hungry investors looking for yield. Sound familiar?

Back when this asset class mushroomed to almost $10 trillion (yes, another T), it was supposed to be different because the Gaussian Copula formula underlying the securitizing and tranching of these mortgages works. See The Formula That Killed Wall Street

But it didn’t this time, as it didn’t anticipate a significant decline in home prices.

We are now building new models using technology to finance invoices. It’s supposed to be different this time because we are clever, we are smarter, we are using artificial intelligence and machine learning algorithms, and can assess the likelihood that invoices will be paid as soon as they are submitted.

For those not familiar with invoice finance, there are three stages where it can be done, each with different risks: invoice submission, when delivery has been verified, and when invoice has been approved by the buyer and scheduled for payment based on payment terms. See: Can Source-to-Pay Networks Go Beyond the Approved Invoice?

Look I am not saying any of the emerging models will fail. Sure, if I have 1,000 invoices submitted to Pacific Gas and they were assessed for a payment, all were paid (most in full), then yes, you can assess that risk. If I am a new supplier, or there is a track record of invoice disputes between the buyer and their sellers, that risk gets harder to assess.

What I am saying about these emerging models are three things:

  1. These models are built without significant credit cycle information.
  2. These models may work for a period of time, but when they don’t, it can become ugly — think back to MBS and the Gaussian Copula formula.
  3. You need more information than just a buyer and their sellers transaction history — that is necessary but not sufficient credit information.

The reality is we are in unchartered waters today, led by central bankers who are increasingly concerned by a liquidity trap and have made qualitative-easing programs worldwide permanent policy.

I’m all for innovative underwriting and bringing automation to the invoice finance market, but models are not bulletproof.

David Gustin runs a research and advisory practice centered on helping financial institutions, vendors and corporations understand the intersection of trade credit, payments and the financial supply chain. This post was written while David worked on a special project with The Interface Financial Group.

Share on Procurious

First Voice

  1. Andres Abumohor:

    Hi David, you are raising an important alert, which is not only true for invoice financing, but by definition, for unprecedented times there is no data that can support back testing exercise, whatever modeling you want to build.
    However, invoice financing has a very short duration portfolio and therefore presents a huge advantage when comparing to any other financial AI model being built.
    Also, I challenge the data sources being used to produce the algorithms today. When it comes to companies, do you have enough real-time data? Or are you just using a huge amount of data? Most of what I have observed around is huge amounts of data from very traditional players based on financial statements, and just some interaction with proprietary platforms… I think alternative data sources is the key around this.

Discuss this:

Your email address will not be published. Required fields are marked *

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