💰 quantitative finance

Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification

This study introduces a novel multi-agent debate framework that simulates the Federal Open Market Committee's deliberative process by modeling interacting LLMs with latent hawkish-dovish beliefs, significantly improving the accuracy and interpretability of monetary policy decision forecasts compared to standard static models.

Original authors: Kaito Takano, Masanori Hirano, Kei Nakagawa

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kaito Takano, Masanori Hirano, Kei Nakagawa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess what a group of 12 people will decide to do next. These people are the "Federal Open Market Committee" (FOMC), the group in charge of setting interest rates in the United States. Their decisions are huge: they can make your mortgage rates go up, your savings account earn more, or even affect the price of your morning coffee.

Usually, when people try to predict what this group will do, they just look at the news and the numbers (like inflation or unemployment) and ask a single computer program, "What do you think?"

The Problem with the Old Way
The authors of this paper say that's like asking one person to guess the outcome of a complex family argument. In reality, the FOMC isn't a single mind; it's a room full of people with different personalities. Some are "Hawks" (they worry a lot about inflation and want to raise rates aggressively), and some are "Doves" (they worry more about the economy slowing down and prefer keeping rates low). They talk, argue, and eventually reach a compromise.

The old computer models missed this "argument" part. They treated the committee like a single robot, ignoring how these different personalities influence each other.

The New Idea: A Digital Debate Club
To fix this, the researchers built a new kind of computer model that acts like a digital debate club.

The Cast of Characters: Instead of one robot, they created a team of 7 different AI "agents." Each agent is programmed with a specific personality, or "belief." Some are programmed to be "Strong Hawks," some "Moderate Doves," and some "Neutrals."
The Inputs: Every agent gets the same homework: a report called the "Beige Book" (which is like a collection of stories from local businesses about how the economy is doing) and some hard numbers (like the unemployment rate).
The First Guess: At the start, every agent looks at the homework and makes a guess on its own. The "Strong Hawk" might guess "Raise Rates," while the "Strong Dove" might guess "Lower Rates."
The Debate: This is the magic part. The agents don't just stick to their first guess. They take turns reading what the other agents said.
- Agent A says: "I think we should raise rates because inflation is high."
- Agent B (the Dove) reads this and thinks: "Hmm, inflation is high, but the local businesses I read about are struggling. Maybe I should soften my stance."
- Agent B updates their guess.
The Consensus: They keep doing this for several rounds. They argue, listen, and adjust their views until they mostly agree on a final decision.

The "Secret Sauce": The Latent Belief
The researchers also added a special feature to make the AI more transparent. They gave each agent a hidden "belief variable." Think of this like a compass inside the agent's brain.

Even though the agent is reading the same news as everyone else, its "compass" points in a different direction. The paper uses math to show that this compass (the belief) is what filters how the agent sees the information. It explains why the Hawk sees a high unemployment number as a reason to keep rates low, while the Dove might see it differently. This makes the AI's decision-making process easier to understand, rather than a "black box."

What Did They Find?
The team tested this system using real historical data from 2000 to 2025. They asked: "Can our digital debate club guess what the real FOMC actually decided?"

The Winner: The system that included the debate and the different personalities was the most accurate.
The Importance of Stories: They found that the "Beige Book" (the stories from local businesses) was actually more important than just the raw numbers. The AI needed the "human stories" to make good guesses.
The Power of Talking: When they turned off the debate feature and just let the agents vote without talking to each other, the system got worse. It showed that the process of arguing and changing minds is crucial for getting the right answer.
The "Hold" Bias: Without the different personalities and the debate, the AI tended to be too safe and just guess "Hold" (do nothing) all the time. The debate helped break this bias and allowed for more accurate predictions of rate hikes or cuts.

In Short
This paper shows that to predict complex financial decisions, you can't just use a smart calculator. You need to simulate a room full of people with different opinions talking it out. By giving AI agents different "personalities" and letting them debate, the model becomes much better at predicting what the real-world central bank will do.

Technical Summary: Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs

Problem Statement

Accurately forecasting Federal Open Market Committee (FOMC) policy decisions (Raise, Hold, or Lower) is critical for financial markets and macroprudential policy. However, existing literature faces two primary limitations:

Neglect of Deliberative Dynamics: Most studies treat the FOMC as a monolithic entity, failing to model the internal deliberative process where members with diverse "hawkish" (tightening) or "dovish" (accommodating) beliefs debate and reach a consensus.
Static Analysis: Prior methods rely on static text analysis or single-model predictions, struggling to cohesively integrate unstructured qualitative texts (e.g., the Beige Book) with quantitative macroeconomic indicators while modeling the evolution of individual beliefs.

This study addresses the challenge of predicting FOMC decisions by structurally simulating the committee's collective decision-making process using Large Language Models (LLMs).

Methodology

1. Multi-Agent Debate Framework

The proposed framework models the FOMC as a system of $n$ autonomous LLM agents (specifically set to $n=7$ to facilitate majority voting). Each agent represents a committee member with a distinct, pre-assigned policy belief profile (e.g., Strong Hawkish, Moderate Dovish, Neutral).

Inputs: Agents receive a combination of qualitative text (Beige Book excerpts) and quantitative data (unemployment rate, inflation rate, historical policy rates).
Iterative Process: The process runs for $T$ $T$ rounds (up to 10).
- Round $t=0$ : Each agent generates an initial prediction ( $z^{(0)}_i$ ) based solely on its internal parameters and the input data.
- Round $t>0$ : Agents observe the predictions and justifications of all other agents from the previous round ( $Z^{(t-1)}$ ). They update their own prediction ( $z^{(t)}_i$ ) by incorporating this social context alongside the original data.
Termination: The debate concludes when all agents reach consensus or after the maximum number of rounds. The final output is the majority vote.

2. Latent Belief Modeling

To enhance interpretability, the authors introduce a discrete latent variable $\theta \in \Theta$ representing each agent's underlying "hawk-dove" stance.

Theoretical Formulation: The decision process is formalized as a Bayesian generative model. The probability of an agent's output is decomposed into the probability of the output given a specific belief, and the posterior probability of that belief given the inputs and peer responses.
Key Assumption: Once a latent belief $\theta$ is fixed, the agent's final label is conditionally independent of the specific input text and data. This implies that beliefs mediate how agents interpret external evidence and peer influence.
Mathematical Structure:
$P(z^{(t+1)}_i | x, v, Z^{(t)}, \phi_i) = \sum_{\theta \in \Theta} P(z^{(t+1)}_i | \theta, \phi_i) P(\theta | x, v, Z^{(t)}, \phi_i)$
Where the posterior belief $P(\theta | \dots)$ is updated based on the likelihood of the observed data and the alignment of peer responses with that belief.

3. Experimental Setup

Dataset: Scheduled FOMC meetings from January 2000 to December 2025. The dataset includes 60 specific time slices (balanced to 15 Raise, 30 Hold, 15 Lower) selected to avoid trivial consecutive trends.
Data Sources:
- Text: "Overall economic activity" or "summary" sentences from the Beige Book.
- Quantitative: Inflation and unemployment rates (FRED), historical policy rates (Bloomberg).
Model: GPT-4o-mini with a temperature of 1 to encourage diverse debate patterns. Structured outputs were enforced to extract labels and justifications.
Ablation Studies: The study compared the full model against variants removing:
1. Beige Book text.
2. Historical policy rate trends.
3. Explicit belief assignments (using a standard majority vote without belief prompts).
4. Multi-round debate (using only Round 0 predictions).

Key Results

Performance Metrics

The full proposed method (Experiment 1) achieved the highest performance among all configurations:

Precision: 0.549
Recall: 0.467
F1-Score: 0.476

Ablation Findings

Importance of Text: Removing the Beige Book (Experiment 2) significantly degraded performance (F1: 0.385), confirming that qualitative regional economic data is crucial for accurate forecasting.
Role of Debate: The "No Debate" baseline (Experiment 6, majority vote of Round 0) yielded an F1 of 0.415. The iterative debate process improved accuracy, demonstrating that social influence and opinion refinement are beneficial.
Impact of Beliefs: Removing explicit belief assignments (Experiment 5) resulted in an F1 of 0.399. More critically, the belief-free model exhibited a strong bias toward predicting "Hold" (280/420 predictions). Introducing diverse beliefs mitigated this bias, leading to more diverse and accurate outcomes.
Historical Rates: Removing historical policy rate trends (Experiment 4) had a negligible impact on accuracy, likely because the dataset was curated to focus on transition points where trends were less predictive.

Behavioral Analysis

The transition matrix revealed that the debate process effectively shifted agents from initial biases toward a consensus. For instance, "Strong Hawkish" agents initially predicted "Raise" in 33 cases but adjusted to 30 after debate, while "Neutral" agents showed significant shifts in their final distributions. The model successfully avoided significant directional errors (e.g., predicting "Raise" when "Lower" was the actual outcome).

Significance and Claims

The paper claims to be the first to apply structured, debate-based multi-agent LLM frameworks specifically to the financial domain of monetary policy. Its primary contributions are:

Structural Simulation: It moves beyond treating central banks as black boxes by explicitly modeling the deliberative, consensus-building nature of the FOMC.
Interpretability: By introducing latent belief variables, the framework provides a theoretical mechanism to understand how individual predispositions (hawkish/dovish) mediate the interpretation of economic data and social influence.
Empirical Validation: The study demonstrates that incorporating institutional features (diverse beliefs and iterative debate) yields superior predictive accuracy compared to standard single-model or static approaches.

The authors maintain a modest stance, acknowledging limitations such as the discrete and small size of the belief space, reliance on a single LLM family, and the restriction to FOMC data. They suggest future work could explore continuous belief spaces and apply the framework to other central banks (e.g., ECB, BOJ).