Amid a push for digital transformation, banks may be tempted to use artificial intelligence and machine learning solutions to "game the system," according to a senior European regulator.
About a third of 60 global banks polled by the Institute for International Finance in 2019 said they are actively using machine learning techniques in models used for credit scoring and as input for regulatory capital calculations, Benoît Cœuré told an Aug. 19 webinar organized by the Peterson Institute for International Economics. Cœuré is a former ECB executive board member who now heads the innovation hub of the Bank for International Settlements.
Financial supervisors have been slower to embrace digitization and, with the COVID-19 pandemic accelerating digital innovation, central banks must attempt to catch up and even stay ahead of the curve, he said.
'Black box' risk
As they are more widely adopted, the new technologies can enhance vigilance in risk monitoring while also improving the resilience and stability of the broader financial system, he said. But their rapid adoption by banks and regulators may also pose risks, he added.
Following the global financial crisis of 2008, the "complexity and opacity of internal models enabled some banks to game the system, resulting in unsatisfactory levels of capital to reflect risks during the financial crisis," Cœuré said. "We also came to realize that both boards and their supervisors had little understanding of the risk parameters being used, contributing to an excessive degree of risk-weighted asset variation."
The current debates around AI, machine learning and the "black box" risk they can create in decision-making are reminiscent of that post-crisis time, Cœuré said.
"In particular, it may be difficult for human users at financial institutions — and regulators — to grasp how outputs and decisions generated by AI and machine learning tools have been formulated and can be explained," he said.
Furthermore, there could be a future situation where banks are aware of the algorithm used by the supervisor in a stress test, for example, he said.
"If you know what machine learning algorithm the supervisor will run on your activities there will be a temptation to game [it]," Cœuré said.
Another challenge of digital transformation for both banks and regulators is the cost of innovation.
Smaller national supervisors will not be able to pay for accelerated digitization and are likely to lag supranational authorities such as the ECB, Cœuré said.
Scale will not be as relevant in financial services because new market entrants that are "nimble enough" and not weighed down by legacy systems may still keep pace with digital innovation in the sector, he said. It is not so much about big versus small but rather who could afford the regulatory technology cost, he noted.
In that respect, large incumbent players that do not have the resources to invest in technology will be most challenged, Cœuré said. Given their low average profitability relative to U.S. peers, big European banks look like the obvious underperformer in terms of digital innovation, according to Cœuré.
This has the potential to increase the existing performance divide between the large U.S. and European players, he noted.
European banks that took longer to recover from the global financial crisis have lagged their counterparts across the Atlantic for almost a decade now. Suffering from a chronic lack of profitability, many of the leading European banks have not been able to rack up investments in technology like the U.S. players.