Comparing the Accuracy of Hybrid Audit Opinion Prediction Models Based on Particle Swarm Optimization (PSO) and Barnacle Optimization (BMO) Metaheuristic

Abstract

This study aims to compare the accuracy of hybrid audit opinion prediction models based on Particle Swarm Optimization (PSO) and Barnacle Optimization (BMO) metaheuristic algorithms. This applied research adopted a descriptive–analytical design. The statistical population included companies listed on the Tehran Stock Exchange between 2015 and 2024, with 138 firms selected through systematic elimination. Data were collected from audited financial statements in the CODAL system and Rahavard Novin software and analyzed using MATLAB 2024 and EViews 13. The multilayer perceptron (MLP) neural network served as the baseline model, while PSO and BMO algorithms optimized its weights and biases. Model performance was assessed using the confusion matrix, Receiver Operating Characteristic (ROC) curve, Area Under the Curve (AUC), correlation coefficient, paired t-test, ANOVA, Friedman, and DeLong tests. All four hypotheses were supported. Optimization of the MLP neural network using PSO and BMO algorithms significantly enhanced prediction accuracy compared with the base model. The BMO algorithm exhibited more robust and stable performance than PSO. The proposed hybrid model combining Random Forest (RF), Extreme Gradient Boosting (XGBoost), and BMO-optimized MLP achieved the highest performance, with 98.9% accuracy, 98.8% precision, and an AUC value of 0.995. The DeLong and Friedman tests confirmed that the BMO-based model significantly outperformed the PSO-based model. Integrating machine learning and metaheuristic optimization substantially improved audit opinion prediction accuracy. The proposed RF–XGBoost–MLP–BMO hybrid framework emerged as the most effective and reliable structure, enhancing auditors’ data-driven decision-making while minimizing human bias and error.