A recent paper from the Financial Stability Board[1]
considers the implications for artificial intelligence (AI) and machine
learning in a number of financial services sectors, including credit risk.
The paper includes a useful section on background and
definitions, and provides a clear reminder that these tools identify patterns
and correlations rather than causality. I suspect that we will need to be
reminded of this distinction more and more, as these tools are being used to
explore complex relationships.
When it comes to credit risk scoring, the FSB is clear that
AI may help to make lending decisions quicker. However, regulators are not
persuaded that AI credit scoring models outperform traditional models – or at
least, “it has not been proved”. For example, a recent paper from Moody’s[2]
compares the performance of their own credit scoring model for corporates
against three machine learning approaches. Moody’s finds that, on average, the
accuracy levels of the four models are comparable, and notes that the key to
enhancing credit scoring models is data.
The FSB notes that the deployment of these AI tools would also
allow access to credit to people or businesses whose creditworthiness cannot be
reliably assessed through traditional credit scoring models. The FSB believes that
this would be a positive development for countries with shallow credit markets (emerging
markets?), though less positive for countries with deep credit markets (developed
markets?). You have been warned…
Regulators are also concerned with the overall auditability
of artificial intelligence models used for credit scoring and the wider impact
on credit risk governance. There is an important dimension here about how the
model is used in business. Is it operating with some human oversight? This is
an important issue for business culture as it forces a consideration of who is ultimately
in control. I suspect that the distinction between retail and commercial
lending in terms of volume of transactions may become important; the volume of retail
transactions might make human oversight more challenging.
Where does that leave the CEO, CFO or CRO of a bank
contemplating the use of AI tools? Here are a few suggestions:
1. Have a shared view of the expected business
outcomes from deploying AI tools.
2. Keep monitoring credit risk exposures and
alignment with risk appetite even more intensively, as the AI tool might have
unintended effects.
3. Focus on the auditability of the AI tool and its
impact on credit risk governance.
If you found this post useful, you may want to subscribe and receive further posts by email – see box on the right hand side of the screen or click here.
[1]
Financial Stability Board, “Artificial
intelligence and machine learning in financial services. Market developments
and financial stability implications”, Nov 2017.
[2]
Bacham, D and J. Y. Zhao, “Machine
Learning: Challenges Lessons and Opportunities in Credit Risk Modelling”,
Moody’s Analytics Risk Perspectives, July 2017.
