An Interpretable Machine Learning Framework for Imbalanced Audit Opinion Prediction

Ha Thanh Nguyen (Ho Chi Minh University of Banking)

Journal of Applied Data Sciences
Vol 7, No 3: September 2026

Audit opinion prediction is important for capital-market monitoring because non-unqualified audit opinions may signal financial distress, reporting uncertainty, and elevated information risk. This study develops an interpretable machine learning framework for predicting non-unqualified audit opinions under severe class imbalance using firm-year observations of Vietnamese listed companies from 2015 to 2024. The proposed framework is based on a Modified Random Forest approach that integrates imbalance-aware sub-sampling, performance-based tree selection, ensemble probability aggregation, and decision-rule extraction. The framework is designed to improve minority-class detection while preserving transparent and economically interpretable risk signals. The empirical analysis compares the proposed framework with several benchmark models, including logistic regression, Support Vector Machine, k-nearest neighbours, Random Forest, XGBoost, Random Forest with synthetic minority oversampling, cost-sensitive Random Forest, and Balanced Random Forest. Predictive performance is evaluated using three validation strategies: a random 70/30 train-test split, an out-of-time split in which observations from 2015-2021 are used for training and observations from 2022-2024 are reserved for testing, and k-fold cross-validation. The proposed framework achieves the strongest overall performance across these settings, with area under the receiver operating characteristic curve values of 0.829, 0.811, and 0.793, respectively, while also improving minority-class recall and F1-score relative to benchmark models. The extracted decision rules indicate that audit opinion modifications are associated with persistent and interacting financial vulnerabilities, particularly prior modified opinions, weak profitability, leverage pressure, limited debt-servicing capacity, liquidity constraints, and asset-structure risk. Methodologically, the study contributes by integrating imbalance-aware ensemble learning with interpretable rule-based analysis. Practically, the framework may serve as a complementary early-warning tool for auditors, regulators, and investors in financial reporting risk assessment within emerging markets.