The Parameter Optimization of Support Vector Machine with Genetic Algorithm in Risk Early Warning Models

Machine Learning algorithms are widely used by lenders in risk early warning models. With Machine Learning, the risk levels of individual and corporate customers are determined at the account and customer level. Lenders want to manage risk by evaluating the payment performance of customer or account with the help of Machine Learning algorithms. Banks, which have an important place among lenders, develop risk early warning models with the help of learning algorithms using customer information. In the development process of risk early warning models, while banks generally use customer information and credit bureau information for the individual segment, they use financial, non-financial and behaviour-based information for the corporate segment. In this study, it is planned to develop a risk early model for customers in corporate service segment. For the customers of corporate service segment, Balance Sheet and Income Statement items were used and the financial ratios were calculated for risk early warning models. In the development of risk early warning models, Mutual Information method was used as a novel feature selection approach and Support Vector Machine method (linear function, radial basis function and sigmoid function) was used as a supervised learning approach. By changing the neighbourhood metric (k), important patterns were discovered with the Mutual Information method in feature selection process. The optimal C and gamma parameters for Support Vector Machine models have been tried to be determined with the Genetic Algorithm, which is among the Meta-Heuristic algorithms. In order to find the optimal metrics in this study, the metric values for all parameters of the SVM model (function specific) have been kept quite wide. In this dataset of corporate service customers, the small neighbourhood metric has been found to have a significant impact on model learning and performance.