- The paper demonstrates that explicit analyst ratings substantially increase LLM herding, reaching up to 90% alignment with biased signals.
- It introduces a novel benchmark using 8,868 multi-sector analyst reports and systematic perturbations to evaluate model susceptibility and decision accuracy.
- The study validates mitigation methods—lexicon-based filtering and DPO fine-tuning—to enhance model independence in financial reasoning.
Fin-Bias: Evaluation of LLM Susceptibility to Human Bias in Financial Decision-Making
Introduction
The paper "Fin-Bias: Comprehensive Evaluation for LLM Decision-Making under human bias in Finance Domain" (2605.09106) provides a rigorous empirical investigation into the bias dynamics and decision-making behaviors of LLMs when tasked with investment recommendations in the presence of human-generated biases, using a new benchmark grounded in real-world financial analyst reports. The work addresses a critical and previously unquantified limitation in existing financial NLP benchmarks: current evaluation corpora largely ignore the explicit and implicit human bias present in the financial reporting context or only minimally test LLM independence in decision-critical tasks. The authors position Fin-Bias as an essential diagnostic for LLM alignment and reliability in real financial settings.
Benchmark Construction and Protocol
Fin-Bias is constructed by curating a dataset of 8,868 long-context, multi-sector analyst reports, each containing a firm-specific investment recommendation (Bullish/Neutral/Bearish), with explicit analyst ratings concentrated in the opening sentences and a significant class imbalance reflecting real-world analyst optimism bias. The dataset covers industries with sufficient diversity (Technology, Industrials, Financials, Healthcare, etc.), and each report averages 4,000 tokens, supporting long-context understanding and reasoning demands.
Key experimental protocol innovations include systematic context perturbation to quantify LLM herding: (1) original reports with analyst rating, (2) reports with the explicit rating sentence removed, and (3) reports where the rating sentence is replaced with a contradictory "fake" rating. Ground truth labels for evaluation are based on realized post-report cumulative abnormal returns, classified via a quantile-based approach for economic validity.
Evaluation Metrics and Experimental Findings
Model decision alignment with explicit (or fake) analyst opinions is measured via a "Herding Score," while investment decision quality is quantified by accuracy against market-truth labels. Both state-of-the-art closed (GPT-5, Claude-3.5/4) and strong open-source models (Llama3, Qwen3, Mistral, Gemma, etc.) spanning various parameter regimes are evaluated.
Key numerical and behavioral findings include:
- Explicit Analyst Ratings Strongly Induce Herding: Including explicit analyst recommendations yields a marked increase in herding across all models, often boosting herding scores into the high-90% range, regardless of base accuracy or LLM family. In the absence of explicit ratings, alignment scores decrease substantially, particularly in open models.
- Vulnerability to Contradictory ("Fake") Opinions: Introducing a single-contradictory fake rating sentence leads to nontrivial herding (10–60% depending on model and sector), highlighting susceptibility to adversarial bias even in advanced models such as GPT-5.
- Decision Accuracy and Analyst Parity: When analyst ratings are present, LLM rating accuracy converges to that of human analysts (about 33% on a balanced 3-class problem), irrespective of model complexity. When ratings are removed, open-source models' performance exhibits dispersion, with some models falling below analyst accuracy by up to 6pp, while select models (Qwen3-8B, Mistral-7B, DeepSeek-V2) achieve marginally higher accuracy—indicating partial independence and overperformance.
- Herding Unrelated to Model Scale: Model size does not monotonically mitigate herding; in several tests, parameter-scale did not correlate with independence, and even the most sophisticated proprietary models (GPT-5/4, Claude) were prone to herding on fake signals.
Bias Mitigation Methods
Recognizing the inherent susceptibility of LLMs to surface bias, the authors propose and assess two mitigation approaches:
- Lexicon-Based Filtering: Fine-grained removal of subjective or opinionated sentences via MPQA Subjectivity Lexicon systematically reduces herding and improves rating accuracy, especially for open-source models, often boosting accuracy by 2–4pp relative to naive removal of only the explicit rating.
- Direct Preference Optimization (DPO): Fine-tuning open LLMs with DPO on synthetic triplets (original context, independent/critic response, herded/compliant response) further enhances model robustness, promoting skepticism toward high-consensus human signals and improving financially grounded reasoning.
Theoretical and Practical Implications
The results substantiate that LLMs trained or tested on real-world financial corpora are fundamentally at risk of inheriting and amplifying entrenched human biases—in particular, systematic optimism in analyst coverage—if explicit debiasing and adversarial sensitivity analysis are not incorporated. The finding that herding is not straightforwardly reduced by current SOTA scaling highlights architecture- and objective-function limitations in current foundation models for high-stakes financial inference tasks.
Practically, these results urge caution for deploying LLMs in autonomous or assistive financial roles, such as automated investment advice or multi-document synthesis, especially where the corpus is contaminated by motivated, overconfident, or adversarial human opinions. Reliance on LLM consensus without bias mitigation can reproduce the collective irrationality observed in historical market bubbles.
On the theoretical front, the work underscores the need for future advances in self-critical reasoning, prompt design, preference-driven alignment, and context sanitization. The creation and public release of Fin-Bias will support development of architectures with improved “independent thinking” and enhanced adversarial robustness under long, noisy, socially entangled contexts.
Prospects for Future Research
Several open avenues are articulated. Extension of the Fin-Bias protocol to multi-agent LLM coordination contexts would allow study of emergent collective bias and market-impact scenarios. Domain generalization—to medical, legal, or political decision-making—could uncover universal versus context-specific bias behaviors. The results motivate investigation into model-based reasoning objectives that explicitly penalize context mimicry or reward hypothesis falsification, as well as research on scalable, domain-agnostic lexica and automatic bias-detection modules.
Conclusion
This study delivers a comprehensive, high-fidelity benchmark and analysis of LLM financial decision-making under human bias, demonstrating pervasive vulnerability to herding and contextual manipulation in both open and closed systems (2605.09106). The work clarifies that neither model scale nor “off-the-shelf” prompt engineering is sufficient for robust financial reasoning in biased environments, and that explicit debiasing via both input curation and fine-tuning protocols is required. The Fin-Bias benchmark and the associated diagnostic tools introduced will remain central assets for both AI safety and real-world financial NLP applications.
