New AI tool can determine necessity of bank bailouts

The bank bailouts that characterized the 2007-2008 financial crisis sometimes had an almost arbitrary feel to the general public as to why some institutions were rescued and some were not. Should a similar crisis happen again, the recent development of an AI-driven algorithm could provide a more precise methodology for policy-makers.

Developed by researchers from University College London, who outlined their process in the journal Nature Communications, the algorithm aims to help governments decide whether or not to bail out a bank in crisis by predicting if the intervention will save money for taxpayers in the long term. They developed the algorithm by using data from the European Banking Authority on a network of 35 European financial institutions judged to be the most important to the global financial system (though it can also be used and calibrated by national banks using detailed proprietary data unavailable to the public).

The algorithm uses a mathematical framework for comparing different bailout strategies in terms of predicted losses to taxpayers, considering factors such as how long the financial crisis is expected to last, the likelihood of each bank defaulting and the effect of a default on other banks in the network, as well as taxpayers' stakes in the banks. Using what's called a Markov Decision Process (effectively, a self-reinforcing mathematical framework for how to react given random or semi-random conditions), researchers then used a bespoke AI algorithm to model the effects of government intervention at any given point in time. The algorithm assesses optimal bailout strategies, comparing no intervention to different types of intervention — that is, varying levels of investment in one bank or many banks — at different time points during a crisis.

Running the algorithm through multiple case studies, researchers found that government bailouts are only optimal when the taxpayers' stake in the banks was greater than a certain critical threshold value, determined via the model. Once the percentage loss goes above this threshold, the optimal policy decision changes. The math itself, which some might be challenged to fully understand, can be reviewed on the paper itself.

The researchers also found that the more distressed a bank or network of banks are, defined in terms of percentage reduction in bank equity, the longer the crisis lasts, and the bigger the bank's exposure to other banks were, the more favorable a bailout tends to be. The researchers also found that, once a bank had received a bailout, the best strategy for taxpayers was if the government continued to invest in that bank to prevent default. This could lead to a lack of incentive for the rescued bank to guard against risk.

"Government bank bailouts are complex decisions that have financial, social and political implications. We believe the AI approach we have developed can be an important tool for governments, helping officials assess specifically financial implications — this means checking if a bailout is in the best interest of taxpayers, or whether it would be better value for money to let the bank fail. Our techniques are freely available for banking authorities to use as tools in their decision-making process," said Dr. Neofytos Rodosthenous, one of the paper's authors.