Forecasting Regional Financial Performance Using Soft-Computing Methods

Abstract

The difficulty in resolving the issues associated with forecasting regional financial performance has spurred the emergence of various applications of soft computing methods to tackle these challenges. This has inspired the development of hybrid models that employ diverse soft computing techniques. In this work, different machine learning methods such as random forest, XGBoost, support vector machines, neural networks, and fuzzy rule-based systems are utilized to improve the prediction of regional financial performance. I propose a novel hybrid method that integrates feature selection, class balancing, and ensemble classifiers in a cost-sensitive prediction scenario. More precisely, the proposed approach aims to develop an accurate decision support system that minimizes the misclassification cost in credit rating classification for subsovereign entities across various countries and world regions. Cost-sensitive learning is employed to adjust the training instances in accordance with the total cost associated with each class, facilitating the prediction of nominal rating classes at a lower misclassification cost. Furthermore, it is demonstrated that combining bagging with decision trees as base learners can mitigate the risk of overfitting, a common issue in individual machine learning methods. To validate the proposed approach, I have conducted experiments using two different types of datasets from Moody's credit rating agency. The results show that the proposed hybrid model surpasses existing forecasting models in terms of misclassification cost and other classification metrics.