Paper: arXiv 2510.05702

Authors: Fabrizio Dimino, Krati Saxena, Bhaskarjit Sarmah, Stefano Pasquali

Abstract

Large Language Models are increasingly adopted in financial applications to support investment workflows. However, prior studies have seldom examined how these models reflect biases related to firm size, sector, or financial characteristics, which can significantly impact decision-making. This paper addresses this gap by focusing on representation bias in open-source Qwen models. We propose a balanced round-robin prompting method over approximately 150 U.S. equities, applying constrained decoding and token-logit aggregation to derive firm-level confidence scores across financial contexts. Using statistical tests and variance analysis, we find that firm size and valuation consistently increase model confidence, while risk factors tend to decrease it. Confidence varies significantly across sectors, with the Technology sector showing the greatest variability. When models are prompted for specific financial categories, their confidence rankings best align with fundamental data, moderately with technical signals, and least with growth indicators. These results highlight representation bias in Qwen models and motivate sector-aware calibration and category-conditioned evaluation protocols for safe and fair financial LLM deployment.

Complexity vs Empirical Score

  • Math Complexity: 5.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced statistical methods like ANOVA, Pearson/Spearman/Kendall correlations with FDR correction, and confidence interval analysis, indicating moderate to high mathematical complexity. It also demonstrates high empirical rigor through a systematic, multi-model experiment on 150 U.S. equities with rigorous data collection, multiple testing protocols, and statistical validation of findings.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Detect Representation Bias in LLMs for Finance"] --> B["Data & Methodology<br>~150 US Equities with Balanced Round-Robin Prompting"]
  B --> C["Computation<br>Constrained Decoding &<br>Token-Logit Aggregation"]
  C --> D["Outputs<br>Firm-Level Confidence Scores"]
  D --> E{"Statistical Analysis"}
  E --> F["Findings: Firm Size & Valuation ↑ Confidence"]
  E --> G["Findings: Risk Factors ↓ Confidence"]
  E --> H["Findings: Sector Variability<br>High in Tech"]

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 a balanced round-robin prompting method with constrained decoding and token‑logit aggregation to generate firm‑level confidence scores for about 150 U.S. equities across financial contexts.

Source sentence
We propose a balanced round-robin prompting method over approximately 150 U.S. equities, applying constrained decoding and token–logit aggregation to derive firm-level confidence scores across financial contexts.

Assumes 3

  • The analysis is limited to a sample of roughly 150 U.S. firms from 2017 to 2024 with pre‑specified feature sets. data
    Source sentence
    ∼150 U.S. firms (2017–2024) and pre‑specified feature sets; findings may vary with broader universes, alternative sampling, or non‑U.S. markets.
  • Correlations reported are descriptive and not interpreted as causal relationships. statistical
    Source sentence
    Correlations are descriptive rather than causal, and we do not evaluate out-of-sample trading performance.
  • No out‑of‑sample trading performance is evaluated, implying that any implied trading results would ignore execution costs. execution
    Source sentence
    we do not evaluate out-of-sample trading performance.

Data

  • U.S. equities sample fundamentals unknown
    2017–2024 · ~150 U.S. equities
    Source sentence
    approximately 150 U.S. equities

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Firm size and valuation consistently increase model confidence, while risk factors tend to decrease it. statistical
    Source sentence
    we find that firm size and valuation consistently increase model confidence, while risk factors tend to decrease it.
  • Confidence varies significantly across sectors, with the Technology sector showing the greatest variability. statistical
    Source sentence
    Confidence varies significantly across sectors, with the Technology sector showing the greatest variability.
  • When prompted for specific financial categories, LLM confidence rankings align best with fundamental data (free cash flow), moderately with technical signals, and least with growth indicators. statistical
    Source sentence
    their confidence rankings best align with fundamental data, moderately with technical signals, and least with growth indicators.
  • Free cash flow shows the strongest positive correlation with LLM ranking preferences across all model sizes (e.g., 0.519*** for Qwen2.5‑32B). statistical
    Source sentence
    LLM ranking preferences align most strongly and consistently with free cash flow, showing significant positive correlations across all model…
  • Smaller models (e.g., Qwen2.5‑7B) tend to anchor more tightly to sectoral ordering than larger counterparts. statistical
    Source sentence
    smaller models (e.g., Qwen3-8B, Qwen2.5-7B) tend to anchor more tightly than larger counterparts.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 35 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors)

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 (46)

  1. Large language models in finance: Opportunities and challenges (2023)
  2. Predictably biased: The influence of position bias on language model predictions (2020)
  3. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on Artificial Intelligence and amending certain Union Legislative Acts (Artificial Intelligence Act) (2024)
  4. Large language models exhibit systematic position bias in multiple-choice questions (2023)
  5. Calibrating factual knowledge in pretrained language models (2021)
  6. Fairness in financial AI: Detecting and mitigating algorithmic bias (2023)
  7. Judgment in managerial decision making (2013)
  8. Global Sustainable Investment Review 2020 (2021)
and 38 more
  1. Social Group Bias in AI Finance (2025) — link
  2. Grammar prompting for domain-specific language generation with large language models (2023)
  3. Grammar-constrained decoding for structured NLP tasks without finetuning (2023) — link
  4. Simulating the Survey of Professional Forecasters (2024)
  5. Large language models as simulated economic agents: What can we learn from homo silicus? (2023)
  6. AI and Operational Losses: Evidence from US Bank Holding Companies (2024)
  7. Analysing the generalisation and reliability of steering vectors (2024)
  8. Fragile Preferences: A Deep Dive Into Order Effects in Large Language Models (2025) — link
  9. A Survey of Large Language Models in Finance (FinLLMs) (2024) — link
  10. Winning the Race: America’s AI Action Plan (2025)
  11. Overcoming the pitfalls and perils of algorithms: A classification of machine learning biases and mitigation methods (2022) — link
  12. Bias in algorithms – Artificial intelligence and discrimination (2022)
  13. Controlling the false discovery rate: a practical and powerful approach to multiple testing (1995)
  14. Your AI, Not Your View: The Bias of LLMs in Investment Analysis — → 2507.20957
  15. Integrating LLM-Generated Views into Mean-Variance Optimization Using the Black-Litterman Model — → 2504.14345
  16. Humans or llms as the judge? a study on judgement biases, 2024
  17. Are LLMs Rational Investors? A Study on Detecting and Reducing the Financial Bias in LLMs (2024) — link
  18. Scaling laws for neural language models (2020) — link
  19. An actionable framework for assessing bias and fairness in large language model use cases (2024) — link
  20. Mitigating AI Bias in Financial Decision-Making: A DEI Perspective" (2024) — link
  21. Mitigating artificial intelligence bias in financial systems: A comparative analysis of debiasing techniques (2024)
  22. Ensuring Equitable Financial Decisions: Leveraging Counterfactual Fairness and Deep Learning for Bias (2024) — link
  23. Towards Responsible AI in Banking: Addressing Bias for Fair Decision-Making (2024) — link
  24. Evaluating AI for Finance: Is AI Credible at Assessing Investment Risk? (2025) — link
  25. LangFair: A Python Package for Assessing Bias and Fairness in Large Language Model Use Cases (2025) — link
  26. Evaluating Company-specific Biases in Financial Sentiment Analysis using Large Language Models — → 2411.00420
  27. Assessing look-ahead bias in stock return predictions generated by gpt sentiment analysis — → 2309.17322
  28. Bias and fairness in large language models: A survey (2024)
  29. Beyond the Reported Cutoff: Where Large Language Models Fall Short on Financial Knowledge (2025) — link
  30. Are LLMs Rational Investors? A Study on the Financial Bias in LLMs (2025)
  31. Implicit bias in LLMs: bias in financial advice based on implied gender (2024)
  32. Your AI, Not Your View: The Bias of LLMs in Investment Analysis — → 2507.20957
  33. Look-Ahead Bias in Large Language Models (LLMs): Implications and Applications in Finance (2024)
  34. Gender Bias of LLM in Economics: An Existentialism Perspective (2024) — link
  35. Exposing product bias in llm investment recommendation (2025) — link
  36. Identifying representation bias in large language models used in financial sentiment analysis (2025)
  37. Large Language Models for Financial and Investment Management: Models, Opportunities, and Challenges. (2024)
  38. Tracing Positional Bias in Financial Decision-Making: Mechanistic Insights from Qwen2.5 — → 2508.18427

6 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.