Adaptive Evidence-Based Multi-Agent Debate for Robust Financial Statement Fraud Screening

Authors

  • Amelia Singh Isenberg School of Management, University of Massachusetts Amherst, USA
  • Derek Das Isenberg School of Management, University of Massachusetts Amherst, USA
  • Emily Li Isenberg School of Management, University of Massachusetts Amherst, USA
  • Nathan Yoon Isenberg School of Management, University of Massachusetts Amherst, USA

DOI:

https://doi.org/10.54691/3xwn1h79

Keywords:

Financial statement fraud, multi-agent debate, evidence grounding, credibility, robust aggregation, accounting and auditing enforcement releases

Abstract

Multi-agent debate has recently been applied to financial statement fraud detection: specialist agents exchange evidence-grounded claims and a reconciled fraud-risk score is issued after a fixed number of rounds. Two questions remain open: when and for how long the agents should debate, and whether debate protects the fused score when one agent hallucinates red flags. We propose A-EMAD, an adaptive evidence-based multi-agent debate in which accounting, peer, temporal and forensic agents decompose their verdicts exactly into claims on nine fraud-risk constructs. Each round cross-examines every claim against the credibility-weighted evidence of the other agents under a proportional burden of proof. We prove that the credibility update is a sup-norm contraction, so the debate converges to a unique fixed point, and we stop each debate adaptively; a disagreement gate decides which filings are debated at all. On the public AAER-labelled panel of Bao et al. (146,045 firm-years), with out-of-time tests on 2003–2008 over four seeds, A-EMAD performs on par with stacking, RUSBoost and a fixed three-round debate (AUC 0.707 vs. 0.706, 0.714 and 0.708; no difference is significant) while using 54% fewer debate rounds than the fixed schedule. When one agent fabricates alert-level red flags, its credibility falls from 0.835 to 0.35, and A-EMAD loses roughly half as much AUC as stacking (p < 10⁻⁴). A median pool is even more robust in AUC on the test years, but it trailed stacking by 0.016 AUC on clean development data and cannot indicate which agent or claim failed.

Downloads

Download data is not yet available.

References

[1] Dyck, A., Morse, A., & Zingales, L. (2010). Who blows the whistle on corporate fraud? The Journal of Finance, 65(6), 2213–2253. https://doi.org/10.1111/j.1540-6261.2010.01614.x DOI: https://doi.org/10.1111/j.1540-6261.2010.01614.x

[2] Beneish, M. D. (1999). The detection of earnings manipulation. Financial Analysts Journal, 55(5), 24–36. https://doi.org/10.2469/faj.v55.n5.2296 DOI: https://doi.org/10.2469/faj.v55.n5.2296

[3] Dechow, P. M., Ge, W., Larson, C. R., & Sloan, R. G. (2011). Predicting material accounting misstatements. Contemporary Accounting Research, 28(1), 17–82. https://doi.org/10.1111/j.1911-3846.2010.01041.x DOI: https://doi.org/10.1111/j.1911-3846.2010.01041.x

[4] Cecchini, M., Aytug, H., Koehler, G. J., & Pathak, P. (2010). Detecting management fraud in public companies. Management Science, 56(7), 1146–1160. https://doi.org/10.1287/mnsc.1100.1174 DOI: https://doi.org/10.1287/mnsc.1100.1174

[5] Bao, Y., Ke, B., Li, B., Yu, Y. J., & Zhang, J. (2020). Detecting accounting fraud in publicly traded U.S. firms using a machine learning approach. Journal of Accounting Research, 58(1), 199–235. https://doi.org/10.1111/1475-679X.12292 DOI: https://doi.org/10.1111/1475-679X.12292

[6] Chen, J., Han, X., Zhang, X., & Yang, Y. (2023). Exact concept-level explanations for early detection of financial misstatements. Eurasia Journal of Science and Technology, 5(2), 34–48. DOI: https://doi.org/10.61784/ejst4162

[7] Bertomeu, J., Cheynel, E., Floyd, E., & Pan, W. (2021). Using machine learning to detect misstatements. Review of Accounting Studies, 26(2), 468–519. https://doi.org/10.1007/s11142-020-09563-8 DOI: https://doi.org/10.1007/s11142-020-09563-8

[8] Craja, P., Kim, A., & Lessmann, S. (2020). Deep learning for detecting financial statement fraud. Decision Support Systems, 139, 113421. https://doi.org/10.1016/j.dss.2020.113421 DOI: https://doi.org/10.1016/j.dss.2020.113421

[9] Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., & Mordatch, I. (2024). Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning (pp. 11733–11763). PMLR.

[10] Liang, T., et al. (2024). Encouraging divergent thinking in large language models through multi-agent debate. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (pp. 17889–17904). https://doi.org/10.18653/v1/2024.emnlp-main.992 DOI: https://doi.org/10.18653/v1/2024.emnlp-main.992

[11] Song, C., Ma, L., Zheng, J., Liao, J., Kuang, H., & Yang, L. (2024). Audit-LLM: Multi-agent collaboration for log-based insider threat detection. arXiv:2408.08902.

[12] Yue, X., Yang, J., & Liu, W. (2026). FraudDebate-Agent: A multi-agent LLM framework with an evidence-based debate mechanism for financial statement fraud detection. Mathematics, 14(15), 2695. https://doi.org/10.3390/math14152695 DOI: https://doi.org/10.3390/math14152695

[13] Public Company Accounting Oversight Board. AS 2401: Consideration of Fraud in a Financial Statement Audit. PCAOB Auditing Standards.

[14] Amayuelas, A., Yang, X., Antoniades, A., Hua, W., Pan, L., & Wang, W. Y. (2024). MultiAgent collaboration attack: Investigating adversarial attacks in large language model collaborations via debate. In Findings of the Association for Computational Linguistics: EMNLP (pp. 6929–6948). https://doi.org/10.18653/v1/2024.findings-emnlp.407 DOI: https://doi.org/10.18653/v1/2024.findings-emnlp.407

[15] Wolpert, D. H. (1992). Stacked generalization. Neural Networks, 5(2), 241–259. https://doi.org/10.1016/S0893-6080(05)80023-1 DOI: https://doi.org/10.1016/S0893-6080(05)80023-1

[16] Järvelin, K., & Kekäläinen, J. (2002). Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems, 20(4), 422–446. https://doi.org/10.1145/582415.582418 DOI: https://doi.org/10.1145/582415.582418

[17] Irving, G., Christiano, P., & Amodei, D. (2018). AI safety via debate. arXiv:1805.00899.

[18] Teng, D. (2025). RSRBench: Reproducibility and performance benchmarking for regional seismic risk workflows. Academic Journal of Architecture and Civil Engineering, 3(1), 40–52. DOI: https://doi.org/10.61784/ajace4020

[19] Rago, A., Toni, F., Aurisicchio, M., & Baroni, P. (2016). Discontinuity-free decision support with quantitative argumentation debates. In Proceedings of the 15th International Conference on Principles of Knowledge Representation and Reasoning (pp. 63–73).

[20] Yin, X., Han, J., & Yu, P. S. (2007). Truth discovery with multiple conflicting information providers on the web. In Proceedings of the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1048–1052). https://doi.org/10.1145/1281192.1281309 DOI: https://doi.org/10.1145/1281192.1281309

[21] Jacobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation, 3(1), 79–87. https://doi.org/10.1162/neco.1991.3.1.79 DOI: https://doi.org/10.1162/neco.1991.3.1.79

[22] Blanchard, P., El Mhamdi, E. M., Guerraoui, R., & Stainer, J. (2017). Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in Neural Information Processing Systems, 30, 119–129.

[23] Ke, G., et al. (2017). LightGBM: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30, 3146–3154.

[24] Ledoit, O., & Wolf, M. (2004). A well-conditioned estimator for large-dimensional covariance matrices. Journal of Multivariate Analysis, 88(2), 365–411. https://doi.org/10.1016/S0047-259X(03)00096-4 DOI: https://doi.org/10.1016/S0047-259X(03)00096-4

[25] Hoberg, G., & Phillips, G. (2016). Text-based network industries and endogenous product differentiation. Journal of Political Economy, 124(5), 1423–1465. https://doi.org/10.1086/688176 DOI: https://doi.org/10.1086/688176

[26] Platt, J. C. (2000). Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In A. J. Smola, P. Bartlett, B. Schölkopf, & D. Schuurmans (Eds.), Advances in Large Margin Classifiers (pp. 61–74). MIT Press.

[27] Lundberg, S. M., et al. (2020). From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence, 2(1), 56–67. https://doi.org/10.1038/s42256-019-0138-9 DOI: https://doi.org/10.1038/s42256-019-0138-9

[28] DeLong, E. R., DeLong, D. M., & Clarke-Pearson, D. L. (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics, 44(3), 837–845. https://doi.org/10.2307/2531595 DOI: https://doi.org/10.2307/2531595

[29] Sun, X., & Xu, W. (2014). Fast implementation of DeLong’s algorithm for comparing the areas under correlated receiver operating characteristic curves. IEEE Signal Processing Letters, 21(11), 1389–1393. https://doi.org/10.1109/LSP.2014.2337313 DOI: https://doi.org/10.1109/LSP.2014.2337313

[30] Zhou, F. (2026). Moment-constrained side-force allocation for wind-assisted merchant ships: Hydrodynamic performance of retractable anti-leeway fins and flap rudders. Innovation and Technology Studies, 3(1), 27–41. DOI: https://doi.org/10.61784/its4027

[31] Qin, Y., Rhee, M., Teng, D., Zhang, C., & Zi, B. (2026). Neuro-symbolic constraint verification for LLM-driven internet finance transaction execution. Scientific Reports. https://doi.org/10.1038/s41598-026-65233-w DOI: https://doi.org/10.1038/s41598-026-65233-w

Downloads

Published

2026-10-30

Issue

Section

Articles

How to Cite

Singh, A., Das, D., Li, E., & Yoon, N. (2026). Adaptive Evidence-Based Multi-Agent Debate for Robust Financial Statement Fraud Screening. Academic Journal of Finance and Accounting, 1(5), 17-29. https://doi.org/10.54691/3xwn1h79