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
Source sentence
approximately 150 U.S. equities
Costs
Not applicable
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.
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 (46)
- Large language models in finance: Opportunities and challenges (2023)
- Predictably biased: The influence of position bias on language model predictions (2020)
- 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)
- Large language models exhibit systematic position bias in multiple-choice questions (2023)
- Calibrating factual knowledge in pretrained language models (2021)
- Fairness in financial AI: Detecting and mitigating algorithmic bias (2023)
- Judgment in managerial decision making (2013)
- Global Sustainable Investment Review 2020 (2021)
and 38 more
- Social Group Bias in AI Finance (2025) — link
- Grammar prompting for domain-specific language generation with large language models (2023)
- Grammar-constrained decoding for structured NLP tasks without finetuning (2023) — link
- Simulating the Survey of Professional Forecasters (2024)
- Large language models as simulated economic agents: What can we learn from homo silicus? (2023)
- AI and Operational Losses: Evidence from US Bank Holding Companies (2024)
- Analysing the generalisation and reliability of steering vectors (2024)
- Fragile Preferences: A Deep Dive Into Order Effects in Large Language Models (2025) — link
- A Survey of Large Language Models in Finance (FinLLMs) (2024) — link
- Winning the Race: America’s AI Action Plan (2025)
- Overcoming the pitfalls and perils of algorithms: A classification of machine learning biases and mitigation methods (2022) — link
- Bias in algorithms – Artificial intelligence and discrimination (2022)
- Controlling the false discovery rate: a practical and powerful approach to multiple testing (1995)
- Your AI, Not Your View: The Bias of LLMs in Investment Analysis — → 2507.20957
- Integrating LLM-Generated Views into Mean-Variance Optimization Using the Black-Litterman Model — → 2504.14345
- Humans or llms as the judge? a study on judgement biases, 2024
- Are LLMs Rational Investors? A Study on Detecting and Reducing the Financial Bias in LLMs (2024) — link
- Scaling laws for neural language models (2020) — link
- An actionable framework for assessing bias and fairness in large language model use cases (2024) — link
- Mitigating AI Bias in Financial Decision-Making: A DEI Perspective" (2024) — link
- Mitigating artificial intelligence bias in financial systems: A comparative analysis of debiasing techniques (2024)
- Ensuring Equitable Financial Decisions: Leveraging Counterfactual Fairness and Deep Learning for Bias (2024) — link
- Towards Responsible AI in Banking: Addressing Bias for Fair Decision-Making (2024) — link
- Evaluating AI for Finance: Is AI Credible at Assessing Investment Risk? (2025) — link
- LangFair: A Python Package for Assessing Bias and Fairness in Large Language Model Use Cases (2025) — link
- Evaluating Company-specific Biases in Financial Sentiment Analysis using Large Language Models — → 2411.00420
- Assessing look-ahead bias in stock return predictions generated by gpt sentiment analysis — → 2309.17322
- Bias and fairness in large language models: A survey (2024)
- Beyond the Reported Cutoff: Where Large Language Models Fall Short on Financial Knowledge (2025) — link
- Are LLMs Rational Investors? A Study on the Financial Bias in LLMs (2025)
- Implicit bias in LLMs: bias in financial advice based on implied gender (2024)
- Your AI, Not Your View: The Bias of LLMs in Investment Analysis — → 2507.20957
- Look-Ahead Bias in Large Language Models (LLMs): Implications and Applications in Finance (2024)
- Gender Bias of LLM in Economics: An Existentialism Perspective (2024) — link
- Exposing product bias in llm investment recommendation (2025) — link
- Identifying representation bias in large language models used in financial sentiment analysis (2025)
- Large Language Models for Financial and Investment Management: Models, Opportunities, and Challenges. (2024)
- 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.