A Majority Voting Mechanism-Based Ensemble Learning Approach for Financial Distress Prediction in Indian Automobile Industry

Financial distress poses a significant risk to companies worldwide, irrespective of their nature or size. It refers to a situation where a company is unable to meet its financial obligations on time, potentially leading to bankruptcy and liquidation. Predicting distress has become a crucial application in business classification, employing both Statistical approaches and Artificial Intelligence techniques. Researchers often compare the prediction performance of different techniques on specific datasets, but no consistent results exist to establish one model as superior to others. Each technique has its own advantages and drawbacks, depending on the dataset. Recent studies suggest that combining multiple classifiers can significantly enhance prediction performance. However, such ensemble methods inherit both the strengths and weaknesses of the constituent classifiers. This study focuses on analyzing and comparing the financial status of Indian automobile manufacturing companies. Data from a sample of 100 automobile companies between 2013 and 2019 were used. A novel Firm-Feature-Wise three-step missing value imputation algorithm was implemented to handle missing financial data effectively. This study evaluates the performance of 11 individual baseline classifiers and all the 11 baseline algorithm’s combinations by using ensemble method. A manual ranking-based approach was used to evaluate the performance of 2047 models. The results of each combination are inputted to hard majority voting mechanism algorithm for predicting a company’s financial distress. Eleven baseline models are trained and assessed, with Gradient Boosting exhibiting the highest accuracy. Hyperparameter tuning is then applied to enhance individual baseline classifier performance. The majority voting mechanism with hyperparameter-tuned baseline classifiers achieve high accuracy. The robustness of the model is tested through k-fold Cross-Validation, demonstrating its generalizability. After fine-tuning the hyperparameters, the experimental investigation yielded an accuracy of 99.52%, surpassing the performance of previous studies. Furthermore, it results in the absence of Type-I errors.