Conventional Machine Learning Methods for Detecting Tax Evasion among Inactive or Insolvent Companies

Previous scientific literature suggests that further research into the tax behaviour of inactive or insolvent companies is required. This paper applied conventional machine learning models rather than advanced ones to maintain explainability and trust in machine learning. One of the challenges of using machine learning to detect tax evasion is the limited availability of private information that can be used as input data. This paper uses public data published by the Romanian Ministry of Finance, the Romanian Trade Register and the Romanian tax authorities. The research questions focus on the preference for supervised versus unsupervised models for tax evasion detection and the statistical significance of the available explanatory variables. Given the significant imbalance in the dataset, this study shows that the performance metrics of classical machine learning models, such as logistic regression and fuzzy c-means, are comparable to those of other models. This paper proposes a two-layer ensemble that is optimised based on an increased weight granted to the recall indicator, with the aim of improving the performance of tax evasion detection. The results show that the business activity code, solvency ratio and company life period are the main explanatory variables of tax evasion in the selected dataset, and that models selected in this research are many times more effective than a random choice. From an economic point of view, this highlights the prolonged survival of inefficient companies, confirms the susceptibility of certain industries to tax evasion, and indicates that financial distress is a driver of tax non-compliance. The contribution of this paper is identifying the main explanatory variables for an automated system for predicting tax evasion among inactive and insolvent companies based on public data. The future emphasis shall be on identifying potential explanatory variables from other sources to improve the performance of robust and explainable machine learning algorithms.