People panic. And even more so when they see other people panic. It could be a rush for the exits when someone spreads a rumour about a fire in a closed hall or a scramble to withdraw money from a bank that is feared to be running out of cash.
Financial instability can be triggered by investor concerns about the economic situation, the actions of other investors who may be prone to panic, or both. It is actually rational to get ahead of the crowd in many such situations, even when one has a calmer view of the underlying reality. Can artificial intelligence (AI) agents do better?
Earlier this year, four economists at the European Central Bank published a paper that goes into this question (Ex Machina: Financial Stability In the Age of Artificial Intelligence by Kartik Anand, Sophia Kazinnik, Agnese Leonello and Ettore Panetti).
As AI plays a bigger role in financial markets, from trading, portfolio advice and risk management to fraud detection, does it make the financial system safer or introduce new types of algorithmic panic? The authors test this by simulating a spiral of mutual fund redemptions that in many ways resembles a classic bank run.
However, instead of human agents deciding whether to sell their mutual fund holdings or continue to hold them, the four economists use two types of AI in their model.
The first are Q-learning agents, while the second are large language models (LLMs). The central insight is that different types of algorithms reacting to the same economic environment produce very different agent behaviour, and thus aggregate outcomes, in terms of financial stability.
Here are the details. Q-learning AI agents learn purely from experience. They try various strategies and learn from what works and what does not. In a financial market, they are more likely to act in unison, so that everyone either stays invested or everyone rushes to redeem mutual fund investments.
There are only two possible types of aggregate behaviour. However, such AI agents are more likely to tilt towards pessimism, making a rush for an exit in the style of a typical bank run possible even if the fundamentals do not justify such a response.
LLM agents reason more in the style of textbook rational investors. They reason about the future like a person would. They do not panic unnecessarily, but because each LLM agent is free to interpret the situation differently based on the structure of its underlying algorithm, these AI agents tend to form different beliefs about what other AI agents will do.
That disagreement means they fail to coordinate. Instead of moving together—either all calm or all panicking—the market gets messy with less predictable partial withdrawals from mutual funds.
Anand, Kazinnik, Leonello and Panetti say that their results “can fundamentally alter how financial fragility emerges in markets.” This has implications for policymakers. “Some AI agents may produce excessive and unnecessary redemptions, while others may generate unpredictable outcomes. As AI becomes more deeply embedded in financial decisions, understanding these differences will be salient for assessing its implications for financial stability.”
The implication is that financial stability in the age of AI may depend on not just the underlying economic situation or the riskiness of strategies, but also the AI architecture that drives the decisions of AI agents.
The authors outline three broad policy lessons from their work.
First, financial regulators need to have a good understanding of what AI tools are being used in trading rooms, given that different types of AI architecture lead to different types of investment behaviour. This may need to be folded into broader financial sector disclosure rules.
Second, traditional risk management tools that regulators use must be bolstered with “mechanisms that respond to the speed and scale at which AI agents operate.”
Third, the scope of financial literacy should be extended to cover how delegating portfolio decisions to AI agents will generate benefits as well as risks. Otherwise, say the authors, there can be the spread of vibe investing, in which “trust in the tool substitutes for understanding how it works.”
One of the earliest lessons in how new technologies can generate new risks to financial stability was learnt from a US stock market crash on 19 October 1987—what is now remembered as Black Monday.
The Dow Jones index of the US lost almost a quarter of its value in a single session of manic trading. Later investigations highlighted how the increasing use of computers had caused a price cascade, as machines programmed to automatically generate stop-loss orders turned a market correction into a crash. That episode led to the introduction of circuit breakers to cool things down.
Artificial intelligence is bound to be increasingly used in financial markets, which will be beneficial even as it creates new risks to financial stability. Anyone with a taste for the darker side of the story would do well to read The Fear Index, a 2011 novel written by Robert Harris about an algorithm that goes rogue in financial markets.
The author is executive director at Artha India Research Advisors.