26 March 2018

Financial Times: Machine learning helps banks cut fraud and prep stress tests

Institutions have been quietly experimenting with the technologies to catch fraudsters, for instance, and for making credit decisions so they can lend more and lose less. At the more technical end, machine learning and AI are becoming important tools for preparing stress-test submissions, which help set the bar on how much capital banks need.

A November report by the Financial Stability Board, a global standard-setter, on the use of AI in financial services noted, for example, how the Australian Securities and Investments Commission was using machine learning to catch rogue financial services professionals through their marketing materials.

The report noted that in the US the Securities and Exchange Commission’s machine learning techniques were “five times better” than random searches for finding “language that merits referral to enforcement”, and that the tool showed “substantial promise” in the financial services industry more generally.

Fraud detection is one of the most important applications for machine learning in finance. Something as seemingly insignificant as a person using a computer in a non-typical way to complete an online form can point to potential wrongdoing.

In a report last summer, rating agency Moody’s said such technology “contributes significantly to credit risk-modelling applications” — which decide whether a borrower will meet loan repayments or not — because “a machine learning model, unconstrained by some of the assumptions of classic statistical models, can yield much better insights that a human analyst could not infer from the data”.

As well as helping banks sidestep fraudulent borrowers or those with poor credit, it enables them to more confidently lend to those who pass the machine learning tool’s risk tests.

Half of the risk managers surveyed by McKinsey for a report published in October said they expected credit decision times to fall by 25 to 50 per cent because of AI and other digital innovations. The consultancy posited that credit losses “may fall by up to 10 per cent” because of the technology.

The FSB report cites an example of a global corporate investment bank using “unsupervised learning” — in which an algorithm is asked to detect patterns in data that have not been previously labelled — to help predict how much it might lose and how much capital it has. In another example an institution used AI to help model the performance of its capital markets business in stress tests.

Banks are reluctant to estimate the financial benefit of deploying machine learning in their risk management functions. But DataArt’s Mr Utkin says the potential cost savings “are huge”.

It is not all good news, however. AI’s lack of “auditability” could present its own risks, the FSB report says, leading to unintended consequences. Actions decided by trading algorithms, for instance, may have negative effects when they interact with markets.

Full article on Financial Times (subscription required)

© Financial Times