Risk-based auditing is crucial for efficient tax management and significantly reduces audit costs and duration. This study aims to design, implement, assess, and compare the accuracy of various predictive models—Logit 1, MPL-ANN 2, SVM 3, LSVM 4, and two DTA 5 (C5.0 and RFA 6), in predicting corporate taxpayers’ risk within the Iranian tax system. The study focuses on a sample of 156 manufacturing companies listed on the Tehran Stock Exchange from 2010 to 2021. Initially, 160 influential variables were identified using meta-synthesis and Delphi techniques, which were then refined to 57 critical variables through four rounds of expert panel discussions. Data were subsequently gathered from the Tehran Stock Exchange website and tax records. The results reveal the varying degrees of accuracy of the models in predicting low, moderate, and high-risk taxpayers. C5.0 and RFA consistently outperformed other models suggesting their superior capability in detecting different risk categories. Specifically, C5.0 ranked highest for high-risk predictions, followed by Logit, RFA, SVM, MLP-ANN, and LSVM. For moderate-risk, the order was C5.0, RFA, SVM, LSVM, MLP-ANN, and Logit. While, for low-risk, RFA led, followed by C5.0, SVM, MLP-ANN, LSVM, and Logit. These findings highlight the importance of selecting tailored models and advocating using a combination of models based on taxpayers’ risk categories in tax audits. The paper concludes followed by recommendations for future research and study limitations.