Predictive modeling of tax compliance risks: A comparative study of machine learning approaches

Abstract

Modern enterprises grapple with complex financial data and multidimensional risk interdependencies in their operations. Machine learning offers transformative potential for tax risk assessment and smart auditing solutions. This research analyzes 3,232 tax records from regional manufacturing and service sectors (2021–2023) to evaluate three predictive models: SVM, XGBoost, and Random Forest. Results demonstrate Random Forest’s superior performance, achieving 92.00% (manufacturing) and 93.39% (service) accuracy – substantially outperforming XGBoost and SVM (85–90%). Key manufacturing risk indicators follow a “high tax-high volatility-high scrutiny” pattern, with tax burden rate (0.129 weight), profit fluctuation (0.100), and audit frequency (0.091) being most predictive. Service sector risks manifest as “volatility-declaration-tax burden” dynamics, where profit volatility (0.142) emerges as the strongest predictor. These findings both validate machine learning’s efficacy in tax analysis and equip regulators with intelligent risk management tools.