A multi-dimensional assessment framework for corporate tax risk based on Bayesian networks

Abstract In the context of increasing regulatory scrutiny and global tax reforms, corporate tax risk is a critical concern for both corporations and authorities. Assessing tax risk is challenging due to complex financial practices and opaque governance, with traditional methods often failing to capture nonlinear interdependencies and hidden risk patterns. To address these limitations, this research proposes a Multi-Dimensional Assessment Framework using a Differentiable Architecture Search–driven Scalable Bayesian Network (DARTS-SBN) for effective tax risk modelling in Python. The model is trained and tested on the Corporate Tax Risk Dataset comprising 1,900 records with financial, governance, audit, and compliance attributes, focusing on ownership, internal controls, tax rates, audit likelihood, offshore transactions, and prior fines. Preprocessing involves Bayesian Expectation-Maximization for missing data, Isolation Forest for outlier mitigation, and Principal Component Analysis (PCA) for extracting uncorrelated features from high-dimensional financial variables. The novelty lies in integrating DARTS with an SBN, enabling continuous relaxation of discrete BN structure search and efficient gradient-based optimization of network topology. The method using Python 11, an approach that automates the discovery of optimal BN architectures representing tax risk dependencies, even in large, high-dimensional datasets, achieving 98.5% accuracy. DARTS-SBN enhances scalability, convergence speed, and accuracy compared to traditional methods. The framework supports interpretable, real-time tax risk analysis and strategic audit targeting, offering substantial utility for regulators and corporate compliance teams.