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

Paper: arXiv 2503.16974

Authors: Julian Junyan Wang, Victor Xiaoqi Wang

Abstract

This study provides the first comprehensive assessment of consistency and reproducibility in Large Language Model (LLM) outputs in finance and accounting research. We evaluate how consistently LLMs produce outputs given identical inputs through extensive experimentation with 50 independent runs across five common tasks: classification, sentiment analysis, summarization, text generation, and prediction. Using three OpenAI models (GPT-3.5-turbo, GPT-4o-mini, and GPT-4o), we generate over 3.4 million outputs from diverse financial source texts and data, covering MD&As, FOMC statements, finance news articles, earnings call transcripts, and financial statements. Our findings reveal substantial but task-dependent consistency, with binary classification and sentiment analysis achieving near-perfect reproducibility, while complex tasks show greater variability. More advanced models do not consistently demonstrate better consistency and reproducibility, with task-specific patterns emerging. LLMs significantly outperform expert human annotators in consistency and maintain high agreement even where human experts significantly disagree. We further find that simple aggregation strategies across 3-5 runs dramatically improve consistency. We also find that aggregation may come with an additional benefit of improved accuracy for sentiment analysis when using newer models. Simulation analysis reveals that despite measurable inconsistency in LLM outputs, downstream statistical inferences remain remarkably robust. These findings address concerns about what we term “G-hacking,” the selective reporting of favorable outcomes from multiple generative AI runs, by demonstrating that such risks are relatively low for finance and accounting tasks.

Complexity vs Empirical Score

• Math Complexity: 2.5/10
• Empirical Rigor: 6.0/10 (capped from 8.8: no out-of-sample test; see penalty flags)
• Quadrant: Street Traders — practical and empirical, lighter on theory

Why this score: This paper presents a highly rigorous and novel empirical evaluation of LLM output consistency and reproducibility across 3.4 million outputs in finance tasks. While mathematical formulations are minimal (focused on standard agreement and correlation metrics), the extensive experimental setup, simulation analysis, and practical findings on ‘G-hacking’ make it a vital contribution to the field.

Research Flowchart

  flowchart TD
    A[Research Goal: Assess consistency and reproducibility of LLMs in finance and accounting tasks] --> B[Data Inputs: MD&As, FOMC statements, news articles, earnings calls, financial statements]
    B --> C[Methodology: 50 independent runs across 5 tasks using GPT-3.5-turbo, GPT-4o-mini, and GPT-4o]
    C --> D[Analysis: Evaluate 3.4M+ outputs, compare with human experts, test aggregation, and simulate statistical impact]
    D --> E[Key Finding 1: High consistency in classification/sentiment; complex tasks show variability]
    D --> F[Key Finding 2: LLMs outperform human experts in consistency; 3-5 run aggregations boost stability]
    D --> G[Key Finding 3: Downstream statistical inferences remain robust, mitigating 'G-hacking' concerns]