Identifying the Factors Affecting Independent Audit Opinions Using Explainable Artificial Intelligence: Evidence from Türkiye

This study aims to identify financial and non-financial variables associated with independent audit opinions by using predictive models interpreted through explainable artificial intelligence (XAI) techniques, thereby improving transparency in audit-opinion prediction. In this context, 2415 firm-year observations from 238 manufacturing firms listed on Borsa Istanbul for the period 2010–2024 were used. In the study, the independent audit opinion was used as the dependent variable, while 18 financial and non-financial attributes were employed as independent variables. Descriptive statistics and a correlation heatmap analysis were conducted. To examine which variables contribute most strongly to audit-opinion predictions across different model specifications, predictive models were developed using logistic regression, decision tree, random forest, XGBoost, LightGBM, and CatBoost. Three complementary interpretability analyses were performed: built-in feature importance, Shapley Additive Explanations (SHAP) global importance, and SHAP beeswarm analysis. The built-in feature importance heatmap and the SHAP feature importance heatmap for all models were presented comparatively. A descriptive cross-model synthesis was then constructed by averaging normalized built-in feature importance and SHAP importance values. According to this descriptive synthesis, the type of audit opinion in the previous year (X17) received the highest average importance score (0.78), followed by asset turnover ratio (X7) with 0.57, operating sector (X18) with 0.54, current ratio (X1) with 0.54, acid-test ratio (X2) with 0.51, and the natural logarithm of total assets (X10) with 0.50. These results indicate that both financial and non-financial variables contribute to model predictions, although their relative importance varies across algorithms and interpretation methods. The findings contribute to the audit-opinion prediction literature by providing a comparative and interpretable assessment of predictive importance across multiple model classes and may support audit-risk assessment and related decision-support processes.