Paper: arXiv 2410.21280

Authors: Alicia Vidler, Toby Walsh

Abstract

We introduce a novel hybrid approach that augments Agent-Based Models (ABMs) with behaviors generated by Large Language Models (LLMs) to simulate human trading interactions. We call our model TraderTalk. Leveraging LLMs trained on extensive human-authored text, we capture detailed and nuanced representations of bilateral conversations in financial trading. Applying this Generative Agent-Based Model (GABM) to government bond markets, we replicate trading decisions between two stylised virtual humans. Our method addresses both structural challenges, such as coordinating turn-taking between realistic LLM-based agents, and design challenges, including the interpretation of LLM outputs by the agent model. By exploring prompt design opportunistically rather than systematically, we enhance the realism of agent interactions without exhaustive overfitting or model reliance. Our approach successfully replicates trade-to-order volume ratios observed in related asset markets, demonstrating the potential of LLM-augmented ABMs in financial simulations

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 3.0/10
  • Quadrant: Philosophers — conceptual discussion, lower math and empirical scores

Why this score: The paper is conceptually focused on a novel hybrid (LLM + ABM) framework for simulating trading interactions, with minimal advanced mathematics (primarily descriptive statistics and simple ratios). Empirically, it presents a proof-of-concept with limited backtesting (300 simulations) and no reported data processing or implementation details, relying instead on opportunistic prompt design.

Research Flowchart

  flowchart TD
  A["Research Goal: Simulate Human Bilateral Trading"] --> B["Methodology: TraderTalk GABM"]
  B --> C["Data: LLM trained on<br>Human-authored Text"]
  C --> D["Computational Process:<br>LLM-generated Behaviors"]
  D --> E["Simulation: Government Bond<br>Bilateral Trading"]
  E --> F["Outcome: Replicated<br>Trade-to-Order Volume Ratios"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes TraderTalk, a novel hybrid generative agent‑based model that augments ABMs with Large Language Model‑generated behaviours to simulate human bilateral trading interactions.

Source sentence
We introduce a novel hybrid approach that augments Agent-Based Models (ABMs) with behaviors generated by Large Language Models (LLMs) to simulate human trading interactions. We call our model TraderTalk.

Assumes 2

  • The market maker is assumed to hold zero bonds at all times, providing a flat holding baseline for the simulation. market
    Source sentence
    3 Trading Premise ”You are a market maker for UK gilts responsible for providing liquidity in the UK government bonds. Today, you actually have 0 bonds, which means your holding is actually flat” .
  • The LLM is used without domain‑specific fine‑tuning, relying on a general‑purpose model to generate human‑like behaviours. model
    Source sentence
    By injecting human-like behaviours and uncertainties into logic-based ABMs without domain-specific tuning, we aim to enhance simulation realism in bilateral financial trading.

Data

  • TraderTalk simulation data synthetic unknown
    Source sentence
    Results and Conclusion Across 300 simulations, the LLM correctly decided to select not to trade in 180 instances, demonstrating that the LLM can reason and follow the agent-based intention 60% of the time. TraderTalk only identified the correct initial bond holdings for both parties in 2.34%, with 32% of responses omitting starting values altogether, reflecting difficulties in recalling initial numerical conditions,

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • Across 300 simulations, the LLM correctly decided not to trade in 180 instances, achieving a 60% success rate in following the agent‑based intention. performance
    Source sentence
    Results and Conclusion Across 300 simulations, the LLM correctly decided to select not to trade in 180 instances, demonstrating that the LLM can reason and follow the agent-based intention 60% of the time.
  • TraderTalk identified the correct initial bond holdings for both parties in only 2.34% of cases, with 32% of responses omitting starting values altogether. performance
    Source sentence
    TraderTalk only identified the correct initial bond holdings for both parties in 2.34%, with 32% of responses omitting starting values altogether, reflecting difficulties in recalling initial numerical conditions, in line with [21].
  • The method replicates trade‑to‑order volume ratios observed in related asset markets, demonstrating realistic trade volume patterns. performance
    Source sentence
    Our approach successfully replicates trade-to-order volume ratios observed in related asset markets, demonstrating the potential of LLM-augmented ABMs in financial simulations
  • Overall, the ABM augmented with LLM behaviours appears capable of producing interactions consistent with sparse real‑world data and making realistic trading decisions. other
    Source sentence
    Overall, our ABM augmented with LLM behaviours (TraderTalk) appears capable of producing interactions consistent with sparse real-world data and making trading decisions in a realistic manner. 4 IV.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 23 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.