Assessing Consistency and Reproducibility in the Outputs of Large Language Models: Evidence Across Diverse Finance and Accounting Tasks

In one sentence

Summary

This paper addresses the critical methodological concern of reproducibility in Large Language Model (LLM) applications within finance and accounting research. As LLMs are increasingly used for text analysis, their inherent stochasticity raises questions about the reliability of research findings. The authors aim to establish baseline expectations for output consistency across common tasks, comparing different model versions and evaluating the impact of output variability on downstream statistical inference. This work fills a gap in the literature by systematically quantifying reproducibility rather than just performance accuracy. The authors conduct an extensive empirical study using three OpenAI models (GPT-3.5-turbo, GPT-4o-mini, and GPT-4o) across five tasks: binary/multi-class classification, sentiment analysis, summarization, text generation, and prediction. They generate over 3.4 million outputs from 50 independent runs for each task, using diverse financial texts including MD&As, FOMC statements, news, and earnings calls. Consistency is measured using metrics like Fleiss’ Kappa, Cohen’s Kappa, and semantic similarity. The study also compares LLM consistency to human expert annotators, tests aggregation strategies (majority voting/averaging), and performs simulation analyses to assess the impact of output variation on regression results. Findings reveal task-dependent consistency: binary classification and sentiment analysis achieve near-perfect reproducibility, while complex tasks like multi-class classification and numerical prediction show greater variability. Advanced models do not consistently outperform predecessors in consistency. LLMs significantly outperform human experts in consistency, maintaining high agreement even where humans disagree. Simple aggregation across 3-5 runs dramatically improves consistency and may improve accuracy. Crucially, simulation analysis shows that despite output variability, downstream statistical inferences remain robust, with negligible impact on regression coefficients and significance tests, suggesting low risk of "G-hacking" in typical finance applications.

Research problem

The paper addresses the lack of comprehensive evidence regarding the consistency and reproducibility of LLM outputs in finance and accounting research. It investigates how stochastic LLM generation affects the reliability of common NLP tasks and whether this variability undermines downstream statistical analysis or enables selective reporting ("G-hacking").

Core contributions

• First comprehensive assessment of LLM consistency across five common finance/accounting tasks using 50 independent runs.
• Benchmarking of three OpenAI models (GPT-3.5-turbo, GPT-4o-mini, GPT-4o) for output stability, revealing that newer models do not always offer better consistency.
• Demonstration that LLMs exhibit higher consistency than human expert annotators, even in ambiguous cases.
• Evidence that simple aggregation strategies (3-5 runs) significantly improve consistency and potentially accuracy.
• Simulation analysis proving that LLM output variability has negligible impact on downstream statistical inference validity.

Method

• Zero-shot prompting for all tasks to test inherent model capabilities without fine-tuning.
• 50 independent runs per task using different random seeds to capture stochasticity.
• Inter-rater agreement metrics: Fleiss’ Kappa, Krippendorff’s Alpha, Cohen’s Kappa.
• Correlation metrics: Intra-class, Pearson, Spearman, and Concordance correlation.
• Variation metrics: Mean Absolute Relative Difference (MARD) for continuous outputs.
• Semantic similarity measured via cosine similarity of jina-embeddings-v3 embeddings.
• Tone analysis using Loughran-McDonald dictionary and FinBERT.
• Simulation analysis of 10 million regressions to assess impact of output variation on inference.

Results & findings

• Binary classification and sentiment analysis achieve near-perfect reproducibility (Fleiss’ Kappa > 0.93-0.94).
• Multi-class classification of FOMC statements shows moderate consistency (Fleiss’ Kappa = 0.86-0.91).
• Numerical predictions show substantial variation in point estimates, especially for less advanced models.
• LLMs outperform human experts in consistency in over 50% of cases, with humans outperforming LLMs in <9% of cases.
• Aggregation of 3-5 runs increases majority class strength from ~86% to 94% (GPT-3.5) and 82% to 91% (GPT-4o-mini).
• Summary length variation decreases from ~8.5% (single run) to ~3% (five aggregated runs).
• 97.35% of statistical inferences regarding length effects are correct, with Type I/II errors at ~1.3%.
• Risk of "G-hacking" (selective reporting) is found to be relatively low for finance/accounting tasks.

Data & experiments

• Three OpenAI models: GPT-3.5-turbo, GPT-4o-mini, GPT-4o.
• Five tasks: binary classification, multi-class classification, sentiment analysis, summarization, text generation, and prediction.
• Diverse financial texts: MD&As, FOMC statements, financial news, earnings call transcripts (presentations and Q&A), and financial statements.
• 50 independent runs per task with different random seeds, generating over 3.4 million outputs.
• Comparison with human expert annotators using the Financial PhraseBank dataset.
• Simulation of 10,000 iterations (10 million regressions) to test robustness of statistical inferences.

Limitations & caveats

• Study focuses only on OpenAI GPT models; results may not generalize to open-source or other proprietary models.
• Zero-shot prompting was used; few-shot or fine-tuned approaches might yield different consistency patterns.
• Consistency metrics vary by task; complex tasks like prediction show higher variability.
• Aggregation increases computational cost, though the paper argues the benefit outweighs the cost.
• The study does not address temporal consistency (performance over time) or model updates.

Curator notes

Discovery status: report_analysis_focus_promoted. Source links: Evaluating Large Language Models (LLMs) in Financial NLP: A Comparative Study on Financial Report Analysis