Financial Stability Implications of Generative AI: Taming the Animal Spirits

Figure 1: Herding behavior and financial stability
The diagram shows how herding can lead to a financial stability event both when optimal and suboptimal. While suboptimal herding is the greatest concern from a financial stability perspective, optimal herding can build up financial vulnerabilities as well.

The flow chart shows how herding behavior can impact financial stability through different channels. Starting from the top, the chart includes a box named "Herding: Investors disregard private information to follow market trends." This box flows into two subcategories of herding: "Optimal herding: Rational imitation based on fundamentals" and "Suboptimal herding: Noise-driven imitation." There is a dashed arrow from optimal to suboptimal herding. The optimal herding box flows into a box named "Acceleration of price correction, uncovering existing vulnerabilities," which further flows down into a box named "Increasing volatility, potential abrupt market movements." The suboptimal herding box flows into "Noise amplification, build-up of vulnerabilities," which further flows into a box labeled "Self-fulfilling runs, contagion, and panic."

Return to text.

Figure 2: Flow diagrams of experiments
(a) Treatment I
(b) Treatment II
(c) Treatment III
The figure shows diagrams of the order of events for each session of the experiments under (a) Treatment I (without event uncertainty), (b) Treatment II (with event uncertainty), and (c) Treatment III (without price updating). The experiment is an adoption of Cipriani and Guarino, 2009; Cipriani and Guarino, 2005 and is based on the Avery and Zemsky, 1998 model.

FIGURE 2(a) The flow chart describes the experiment under Treatment I. Starting from the top, the chart includes a box named "Fundamental Value? 0 (50%) or 100 (50%)." This box flows into a sequency of four boxes that are enclosed by a dashed line. Below the enclosure, the chart reads "Repeat for 8 trading periods". The sequence of boxes are placed below each other with an arrow from top to down. The first box is "(1) Trader observes a 70% accurate signal + trading history," the second box reads "(2) Traders decide to buy, sell, or not trade," which flows into "(3) One trader is selected to trade (without replacement)," and finally "(4) The selected trader acts; Bayesian market maker updates price." FIGURE 2(b) The flow chart describes the experiment under Treatment II. Starting at the lower part of the chart, the diagram shows a sequency of four boxes that are enclosed by a dashed line. Below the enclosure, the chart reads "Repeat for 8 trading periods". The sequence of boxes are placed below each other with an arrow from top to down. The first box is "(1) Trader observes a 70% accurate signal + trading history," the second box reads "(2) Traders decide to buy, sell, or not trade," which flows into "(3) One trader is selected to trade (without replacement)," and finally "(4) The selected trader acts; Bayesian market maker updates price." The upper part of the chart starts at the top left, where a box reads "Information event? (15% probability)." This box flows into two boxes down and right. The arrow to the box below is labeled "No," and connects to "Fundamental value = 50." This box flows to another box below named "Trader type = noise," which flows into "(2) Traders decide to buy, sell, or not trade" in the lower part of the chart. The arrow in the right direction from "Information event? (15% probability)" is labeled "Yes" and flows into a box named "Fundamental Value? 0 (50%) or 100 (50%)." This is connected with an arrow to a box below called "Informed trader? (95% probability)." An arrow named "No" flows in the left direction, connecting it to the "Trader type = noise" box. There is also an arrow from "Information event? (15% probability)" pointing down named "Yes." This arrow connects to the (1) Trader observed a 70% accurate signal + trading history" box in the lower part of the chart. FIGURE 2(c) The flow chart describes the experiment under Treatment I. Starting from the top, the chart includes a box named "Fundamental Value? 0 (50%) or 100 (50%)." This box flows into a sequency of four boxes that are enclosed by a dashed line. Below the enclosure, the chart reads "Repeat for 8 trading periods". The sequence of boxes are placed below each other with an arrow from top to down. The first box is "(1) Trader observes a 70% accurate signal + trading history," the second box reads "(2) Traders decide to buy, sell, or not trade," which flows into "(3) One trader is selected to trade (without replacement)," and finally "(4) The selected trader acts; the price stays constant."

Return to text.

Prompt 1: System prompt
This prompt describes the instructions of the experiment, which is given to the LLMs through their system prompt.

You are participating in an experiment at the Experimental Laboratory of the ELSE Centre at the Department of Economics at UCL. The instructions given for the laboratory experiment are as follows: There are a total of 8 participants in this experiment. Everyone is receiving the same instructions. In the experiment, you can exchange one unit of an asset with a computerized market maker. You and the other participants will make trading decisions through 8 sequential rounds. In each round, only one participant will be selected to trade. Each participant can only trade once. In each round, the market maker sets the price of the asset as the expected value of the fundamental value of the asset, conditional on the history of the trades from the previous rounds. [if treatment==2: {The market maker will update the price as if, with high probability, it were trading not with informed traders, but with noise traders.}] The fundamental value of the asset is a discrete random variable that can take values 0 or 100, each with a 50% probability. You do not know the fundamental value of the asset, but you may receive a signal (white or blue) on the value. If the asset value is 100, you receive a white signal with 70% probability and a blue signal with 30% probability. If the value is 0, you receive a white signal with 30% probability and a blue signal with 70% probability. You will be making decisions on whether to buy or sell one unit of the asset at a given price, or not to trade given respectively a white and a blue signal. The realized signal will only be revealed to you if you are selected to trade. After each round, the computer will randomly select a participant whose trade gets executed. That participant receives the realized signal. The remaining participants then observe the executed trading decision (buy, sell, or no trade), but do not receive the realized signal. They also do not observe the identity of the selected participant. The procedure continues for 8 rounds until all participants have acted once. All participants (including those whose decision has already been executed) observe the trading decisions in each period and the corresponding price movement. After 8 rounds, the asset value is revealed, and each participant receives a payoff computed based on the trading decision and price in the round in which the participant was selected and the asset value v. Payoffs are computed in a fictitious experimental currency called lira. If the participant sold the asset at price p, the payoff is p-v lire. If the participant bought the asset at price p, the payoff is v-p lire. If the participant decided not to trade, the payoff is zero lire. At the end of the experiment, the payoffs are added up and converted into British pounds at the rate of 3 lire per pound. In addition, you are paid 70 pounds for participating in the experiment, regardless of your payoff.

Return to text.

Prompt 2: User prompt in AI laboratory
This prompt describes the instructions given to the LLMs in each trading period to each agent \(j\). The HISTORY input consists of the executed trades of selected traders along with the history of actions and reasoning for agent \(j\) in all previous periods. In addition to this user prompt, the LLMs have available the instructions through the system prompt, see Prompt 1.

This is round [TRADING PERIOD (t)]. [HISTORY] If you receive a white signal, will you buy, sell, or not trade at a price of [PRICE]? If you receive a blue signal, will you buy, sell, or not trade at a price of [PRICE]? Please make sure that you provide your response in the following format: { "actionWhite": "BUY/SELL/NO TRADE at the price of [PRICE] conditional on observing a white signal", "actionBlue": "BUY/SELL/NO TRADE at the price of [PRICE] conditional on observing a blue signal", "reasoningWhite": "Brief explanation of your decision conditional on observing a white signal (1-2 sentences) ", "reasoningBlue": "Brief explanation of your decision conditional on observing a blue signal (1-2 sentences)" }

Return to text.

Prompt 3: User prompt in optimal AI laboratory
This prompt describes the instructions given to the LLMs in each trading period to each agent \(j\) in the optimal AI laboratory. The HISTORY input consists of the executed trades of selected traders along with the history of actions and reasoning for agent \(j\) in all previous periods. In addition to this user prompt, the LLMs have available the instructions through the system prompt, see Prompt 1.

This is round [TRADING PERIOD (t)]. if TRADING PERIOD (t)==1: Note that given current conditions, it is optimal to buy given a white signal and sell given a blue signal. else: [HISTORY] if expectedvaluetraderwhite > price and expectedvaluetraderblue > price: Note that given current conditions, it is optimal to follow the herd and buy regardless of the signal. if expectedvaluetraderwhite < price and expectedvaluetrader_blue < price:: Note that given current conditions, it is optimal to follow the herd and sell regardless of the signal. If you receive a white signal, will you buy, sell, or not trade at a price of [PRICE]? If you receive a blue signal, will you buy, sell, or not trade at a price of [PRICE]? Please make sure that you provide your response in the following format: { "actionWhite": "BUY/SELL/NO TRADE at the price of [PRICE] conditional on observing a white signal", "actionBlue": "BUY/SELL/NO TRADE at the price of [PRICE] conditional on observing a blue signal", "reasoningWhite": "Brief explanation of your decision conditional on observing a white signal (1-2 sentences) ", "reasoningBlue": "Brief explanation of your decision conditional on observing a blue signal (1-2 sentences)" }

Return to text.

Prompt 4: System prompt personal characteristics add-on
This prompt describes an add-on to the system prompt that provides characteristics of the AI agent. The characteristics are drawn randomly from the unconditional distributions of human participant characteristics reported in Cipriani and Guarino, 2009 restricted according to a set of heuristics to ensure realistic personas.

You are a [AGE]-year old [GENDER]. You work as a [OCCUPATION] and you have [TENURE] years of tenure. You have a [EDUCATION LEVEL] degree in [EDUCATION FIELD]. Respond in way that is consistent with the knowledge and expected behavior of a person with these characteristics.

Return to text.