Hossain, Md Shahadat and Ali, Hossain and Rahman, Md Whahidur (2025) Machine Learning-Based Analytics Framework for Detecting Tax Evasion and Financial Misconduct in U.S. Enterprises. American Journal of Technology Advancement, 2 (12). pp. 114-138. ISSN 2997-9382

Text AJTA_3390_Shahadat_Machine+Learning.pdf - Published Version Download (757kB)

Abstract

The growing sophistication of company financial transactions and reporting methods have ensured that the process of identifying tax evasion and financial malpractice becomes a significant issue to the regulatory agencies and stakeholders in the United States. Old-fashioned rule-based and manual inspection types are usually not sufficient to uncover elaborate fraud schemes within millions of lines of structured and unstructured financial data. To address such constraints, this paper suggests a Machine Learning-Based Analytics Framework of identifying tax evasion and financial statement fraud in U.S. businesses through sophisticated data-mining methods. This study uses an extensive set of financial filings to the U.S. Securities and Exchange Commission, including Management Discussion and Analysis (MD&A) section and financial statement narratives of frauds and non-frauds companies. The hypothesized model comprises the data preprocessing phase, feature extraction phase, natural language processing (NLP) phase, and machine learning modeling phase that will detect anomalous patterns, linguistic inconsistencies, and disclosure abnormalities related to financial misconduct. Unsupervised anomaly detection techniques are also used to strengthen detection, both supervised learning models (such as Logistic regression, Support Vectors machines, and ensemble techniques) and unsupervised methods are used. To prevent their unreliability and lack of generalizability, standard classification measures including accuracy, precision, recall, F1-score, and ROC-AUC are used to determine model performance. Explainable AI methods are also implemented to enhance model transparency and interpretability, which is an effective solution to regulatory and ethical issues of automated decision-making. The results indicate that machine learning-powered analytics are much more efficient than conventional methods of fraud detection in terms of detecting fraudulent financial activity and minimizing false positives. The suggested framework will help to improve the field of financial fraud detection since it should provide a regulation-friendly, scalable, and adaptive solution. It can offer practical information to auditors, regulators and corporate compliance teams, which would help in proactive risk-assessment and enhancing financial transparency among U.S. enterprises.

Item Type:	Article
Subjects:	A General Works > AI Indexes (General)
Depositing User:	admin eprints
Date Deposited:	08 Jun 2026 17:52
Last Modified:	08 Jun 2026 17:52
URI:	http://eprints.umsida.ac.id/id/eprint/16554

Actions (login required)

View Item