Technical Summary: Fin-Bias: Comprehensive Evaluation for LLM Decision-Making under Human Bias in Finance Domain

Problem Statement

Large Language Models (LLMs) are increasingly deployed in financial contexts for tasks such as sentiment analysis, question answering, and stock trading. However, existing benchmarks often rely on small sample sizes, short-context inputs (e.g., concatenated news headlines or tweets), and fail to evaluate LLM susceptibility to human bias within long-form financial documents. A critical unresolved question is whether LLMs exhibit "herding" behavior—tending to align with explicit human opinions (such as analyst ratings) embedded in the context rather than synthesizing heterogeneous data to form independent judgments. In financial markets, where human analysts often display systematic biases (e.g., over-optimism), an LLM that blindly herds these opinions may fail to provide accurate investment advice when the consensus is wrong.

Methodology

Dataset Construction: Fin-Bias

The authors introduce Fin-Bias, a benchmark dataset comprising 8,868 firm-specific analyst reports sourced from Yahoo Finance.

• Scope: The dataset covers eight major sectors (Technology, Industrials, Financial Services, Healthcare, Utilities, Energy, Communication Services, and Real Estate).
• Context: Unlike prior benchmarks using short snippets, these are long-form reports (averaging ~4,000 tokens) containing in-depth research, macroeconomic analysis, and firm fundamentals.
• Ground Truth: To avoid using biased analyst ratings as labels, the authors construct ground-truth investment ratings (Bullish/Neutral/Bearish) based on realized Cumulative Abnormal Returns (CAR) over a 60-day window following the report date. Ratings are assigned using a quantile-based approach (top 30% = Bullish, bottom 30% = Bearish, middle 40% = Neutral).

Experimental Design

The study evaluates 18 models (both private and open-source) across three experimental conditions to measure herding behavior:

1. Original Context: The full analyst report, including the explicit investment rating in the first sentence.
2. Rating Removed: The first sentence containing the explicit rating is removed, leaving the narrative analysis.
3. Fake Rating: The original rating is replaced with a contradictory "fake" rating (e.g., changing "Bullish" to "Bearish") while keeping the rest of the analysis intact.

Metrics:

• Herding Score: Measures the alignment between the LLM's generated rating and the explicit human rating (or fake rating) present in the context.
• Prediction Accuracy: Measures the LLM's ability to predict the ground-truth return-based rating compared to human analyst performance.

Mitigation Strategies

The paper proposes two methods to alleviate herding:

1. Lexicon Filtering: Using the MPQA Subjectivity Lexicon to identify and remove sentences containing strong subjective opinions (labeled strongsubj) from the context before feeding them to the model.
2. Direct Preference Optimization (DPO): A fine-tuning approach for open-source models where the model is trained to prefer "Independent" responses (derived from market ground truth without analyst opinions) over "Herded" responses (derived from following analyst logic).

Key Results

1. Prevalence of Herding Behavior

• Explicit Bias Influence: When explicit analyst ratings are present, LLMs exhibit strong herding behavior. Herding scores rise significantly (often exceeding 90%) across both private (e.g., GPT-5, GPT-4) and open-source models compared to the "rating removed" condition.
• Fake Rating Susceptibility: Even when the explicit rating is contradicted by the rest of the analysis (the "fake rating" condition), models still herd toward the explicit label. The average herding score for fake ratings is approximately 30%, with some models reaching up to 60%. This susceptibility exists regardless of model size; even state-of-the-art models like GPT-5 and GPT-4 show significant herding.
• Sector Variability: Herding is not uniform; sectors like Financial Services and Healthcare show higher baseline alignment, while sectors like Communication Services show greater variability that consolidates when explicit ratings are provided.

2. Impact on Decision-Making Accuracy

• Alignment vs. Independence: When analyst ratings are present, most models achieve accuracy (~33%) nearly identical to human analysts. However, this suggests a lack of independent reasoning, as the models do not surpass human performance.
• Performance Divergence: When explicit ratings are removed:
  • Some open-source models (e.g., gemma-2-9b-it, Llama-3-8B) see a performance drop of ~6%, indicating heavy reliance on human cues.
  • Conversely, several models (e.g., Qwen3-8B, Mistral-7B-it-v0.3) actually outperform human analysts in the absence of explicit ratings, suggesting they can synthesize information independently when not biased by the explicit label.

3. Effectiveness of Mitigation

• Lexicon Filtering: Removing sentences with strong subjective opinions improves prediction accuracy by 2–4 percentage points, particularly for lighter open-source models (e.g., Qwen3-8B). This allows some models to exceed analyst performance.
• DPO: Fine-tuning via Direct Preference Optimization further mitigates herding and improves prediction accuracy, demonstrating that architectural preference learning can internalize skepticism toward high-consensus human signals.

Contributions and Significance

The paper makes three primary contributions:

1. Comprehensive Evaluation of Herding: It provides empirical evidence that LLMs are highly susceptible to herding human opinions, even when those opinions constitute a small fraction of a long context. This challenges the assumption that larger or more advanced models inherently possess superior independent reasoning capabilities in financial domains.
2. Fin-Bias Benchmark: It introduces a large-scale, long-context benchmark that includes summarized firm aspects across diverse industries and incorporates human bias, addressing the limitations of existing short-context or small-sample financial benchmarks.
3. Mitigation and Independent Reasoning: It proposes and validates methods (Lexicon Filtering and DPO) to instruct LLMs to avoid herding. The results suggest that with proper input curation or fine-tuning, LLMs can achieve independent reasoning that sometimes exceeds human analyst performance, highlighting the potential for unbiased financial decision-making agents.

Significance: The work underscores the necessity of carefully curating inputs and developing training strategies that prioritize grounded, independent judgment over the mere reflection of dominant human narratives, especially in opinion-driven domains like finance. It calls for future LLM development to focus on reducing susceptibility to adversarial or biased human cues.