Paper: arXiv 2610.01897

Authors: Victoria Ruojie Li, Arka Prava Bandyopadhyay

Abstract

We study whether model diversity survives selection into trading. In synthetic markets with a fixed mixture of three language-model families, news presentation changes their representation among submitted orders. At the announcement round, Qwen’s share of submitted orders shifts by 48 percentage points in the financing event, with little change in net order counts. In the workforce-reduction event, Mistral’s share shifts by 40 percentage points while net counts reverse sign. Homogeneous populations remove opposing flow when their active decisions share a direction. An analytical decomposition shows why selection can improve or worsen price accuracy even at unchanged aggregate demand sensitivity. The evidence concerns presentation bundles and submitted flow; cleaner replication and a known-value validation are specified prospectively.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a novel approach to studying model diversity in trading using synthetic markets and LLM agents. It demonstrates strong empirical rigor through controlled experiments and statistical analysis of order flow, while also providing an analytical decomposition for price accuracy. The methodology is clearly outlined, though full reproducibility would benefit from direct code/data links.

Research Flowchart

  flowchart TD
    A["Research Goal: Model Diversity & Selection into Trading?"] --> B{"Methodology: Synthetic Markets & News Presentation"};
    B --> C{"Data/Inputs: 3 LLM Families, News Events (Financing, Workforce Reduction)"};
    C --> D["Computational Processes: Analyze Order Flow Shifts, Net Order Counts, Analytical Decomposition"];
    D --> E["Key Findings: News Presentation Alters Model Representation in Order Flow; Selection Can Impact Price Accuracy (Positively/Negatively) Independently of Aggregate Demand Sensitivity"];

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 contributes a measured distinction between installed model diversity and active representation of models in trading, rather than discovering heterogeneous LLM behavior.

Source sentence
Our contribution is a measured distinction between installed and active model composition, not the discovery of heterogeneous LLM behavior.

Assumes 6

  • Agents’ actions are independent conditional on a fixed cell and a fixed population assignment, with buy, sell, and hold probabilities summing to one. model
    Source sentence
    Suppose that, conditional on a fixed cell and a fixed population assignment, agents’ actions are independent, with buy, sell, and hold probabilities bi, si, hi, summing to one.
  • Direction exchangeability among active agents conditional on seed and round is assumed for the null hypothesis. statistical
    Source sentence
    The null assumes direction exchangeability among active agents conditional on seed and round.
  • The event round is exactly t = 5 or t = 12, with t = 5 used for participation and rationale counts. other
    Source sentence
    The event round is exactly t = 5 or t = 12, 5 used for participation and rationale counts.
  • The event window comprises the event round and the following two rounds for co‑action and homogeneous‑flow analyses. other
    Source sentence
    The event window comprises that round and the following two rounds, used for the stated co-action and homogeneous-flow analyses.
Show 2 more
  • Background liquidity is thick and constrains the price response, implying elastic liquidity supply. market
    Source sentence
    The exact prompt and recorded treatment texts are reproduced in Appendix D. The thick background liquidity constrains the price response.
  • Agents face disclosed cash, inventory, fees, a five‑share cap and a mean‑variance objective, and utility regret is measured relative to the best feasible decision. execution
    Source sentence
    Agents face disclosed cash, inventory, fees, a five-share cap and a stated mean–variance objective. For a provider with slope 1/λ, a signed order q at pre-order midpoint m costs C(q; m) = mq + q2 2λ + k|q|, m+ = m + q/λ.

Data

  • Synthetic market simulations with three language‑model families synthetic unknown
    Source sentence
    In synthetic markets with a fixed mixture of three language-model families, news presentation changes their representation among submitted orders.
  • Audited decision logs (115,500 decisions) synthetic unknown
    Source sentence
    No JSON parse failures occur in the 115,500 decisions in that audited set.
  • News text bundles (Design A factual, layoff factual, explanatory, etc.) text-news unknown
    Source sentence
    40 News 06 Used in: Design A factual, t = 5. News 07 Used in: Layoff factual, cash-first, t = 5.

Costs

Not applicable The paper reports no trading or execution result; the cost the extractor found is a model parameter.

The paper-text extraction found a cost parameter inside a model (Order cost includes a linear term, quadratic impact term (q^2/(2λ)) and a fee proportional to |q| (k|q|).); the classifier found no reported trading result, so costs are not applicable.

Finds 5

  • In the financing announcement, Qwen's share of submitted orders shifts by 48 percentage points, with little change in net order counts. performance
    Source sentence
    At the announcement round, Qwen's share of submitted orders shifts by 48 percentage points in the financing event, with little change in net order counts.
  • In the layoff event, Mistral's share shifts by 40 percentage points and net counts reverse sign. performance
    Source sentence
    In the workforce-reduction event, Mistral's share shifts by 40 percentage points while net counts reverse sign.
  • Homogeneous model populations remove opposing flow when the active family supplies an extreme direction. theoretical
    Source sentence
    Homogeneous populations remove opposing flow when their active decisions share a direction.
  • Selection can improve or worsen price accuracy even when aggregate demand sensitivity is unchanged. theoretical
    Source sentence
    An analytical decomposition shows why selection can improve or worsen price accuracy even at unchanged aggregate demand sensitivity.
  • Presentation bundles change the family composition of submitted orders in both synthetic announcement settings. performance
    Source sentence
    In two synthetic announcement settings, presentation bundles change the family composition of submitted orders. The layoff contrast moves both active representation and aggregate direction; the financing contrast shows that large representation changes can coexist with little change in signed order counts.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 13 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.

Cited References (17)

  1. When Transparency Improves, Must Prices Reflect Fundamentals Better? (2018) — link
  2. Generative AI enhances individual creativity but reduces the collective diversity of novel content (2024) — link
  3. The Causal Impact of Media in Financial Markets (2011) — link
  4. Information Diversity and Complementarities in Trading and Information Acquisition (2015) — link
  5. Empirical Asset Pricing via Machine Learning (2020) — link
  6. Does Algorithmic Trading Improve Liquidity? (2011) — link
  7. Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models — → 2304.07619
  8. All the News That’s Fit to Reprint: Do Investors React to Stale Information? (2011) — link
and 9 more
  1. Assessing Look-Ahead Bias in Stock Return Predictions Generated by GPT Sentiment Analysis — → 2309.17322
  2. Can Large Language Models Trade? Testing Financial Theories with LLM Agents in Market Simulations — → 2504.10789
  3. Machine Forecast Disagreement (2026) — link
  4. Does Generative AI Facilitate Investor Trading? Early Evidence from ChatGPT Outages (2025) — link
  5. AI-Powered Trading, Algorithmic Collusion, and Price Efficiency (2025) — link
  6. LLM Agents Do Not Replicate Human Market Traders: Evidence From Experimental Finance — → 2502.15800
  7. Market Signal Injection: Adversarial Context Manipulation of LLM Pricing Agents (2026) — link
  8. Provider-Dependent Behavioral Asymmetries in LLM Trading Agents: Evidence from the Brazilian Equity Market (2026) — link
  9. Estimating means of bounded random variables by betting (2020) — link

4 of 17 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.