| Summary: | Credit scoring is an important tool used by financial institutions to correctly identify<br />defaulters and non-defaulters. Support Vector Machines (SVM) and Random Forest (RF) are the<br />Artificial Intelligence techniques that have been attracting interest due to their flexibility to account<br />for various data patterns. Both are black-box models which are sensitive to hyperparameter settings.<br />Feature selection can be performed on SVM to enable explanation with the reduced features,<br />whereas feature importance computed by RF can be used for model explanation. The benefits<br />of accuracy and interpretation allow for significant improvement in the area of credit risk and<br />credit scoring. This paper proposes the use of Harmony Search (HS), to form a hybrid HS-SVM to<br />perform feature selection and hyperparameter tuning simultaneously, and a hybrid HS-RF to tune<br />the hyperparameters. A Modified HS (MHS) is also proposed with the main objective to achieve<br />comparable results as the standard HS with a shorter computational time. MHS consists of four<br />main modifications in the standard HS: (i) Elitism selection during memory consideration instead<br />of random selection, (ii) dynamic exploration and exploitation operators in place of the original<br />static operators, (iii) a self-adjusted bandwidth operator, and (iv) inclusion of additional termination<br />criteria to reach faster convergence. Along with parallel computing, MHS effectively reduces the<br />computational time of the proposed hybrid models. The proposed hybrid models are compared<br />with standard statistical models across three different datasets commonly used in credit scoring<br />studies. The computational results show that MHS-RF is most robust in terms of model performance,<br />model explainability and computational time.
|
|---|
