Automated Machine Learning for Predicting Corporate Tax Non-Compliance

This study aims to develop a tool for predicting corporate tax non-compliance through automated machine learning (AutoML). Recent advancements in machine learning (ML), including techniques such as artificial neural networks, decision trees, support vector machines, random forests, and ensemble methods, have improved the ability to detect tax evasion or tax non-compliance by identifying patterns that traditional methods, like logistic regression, may overlook. However, these methods often need expert knowledge and computational demand. AutoML addresses these limitations by automating the model development process, reducing the need for expert intervention, and making advanced detection modelling more accessible and cost-effective. This study is one of the first to apply AutoML in predicting tax compliance and solving multiclass categorization problems, allowing for the prioritization of different levels of tax compliance. For this purpose, financial and qualitative data from 1,842 small and medium-sized joint-stock enterprises are collected. The findings demonstrate that the H2O-AutoML tool achieves high accuracy and AUC-ROC (both over 90%) in predicting tax non-compliance, performing well even with imbalanced data. The study contributes to the literature by showcasing the practical advantages of implementing AutoML in tax compliance prediction, particularly in improving model accuracy and efficiency. The study's primary limitation lies in its reliance on specific algorithms and data samples, which may only generalize across some contexts or industries. Future research should focus on validating these findings across various sectors and geographic regions to assess the model's generalizability.