Systemic financial risk prediction using least squares support vector machines

The systemic financial risk prediction problem has become a focus in the field of finance. This work applies a novel machine learning technique, that is, least squares support vector machines (LSSVM), to predict the systemic financial risk. To serve this purpose, the paper selects financial risk indicators of China from January 2006 to December 2016, and utilizes unit root test, principal component analysis (PCA) and self-exciting threshold autoregressive (SETAR) methods for data preprocessing. Furthermore, particle swarm optimization (PSO) has been used for parameters optimization of LSSVM by comparison with grid search (GS), and genetic algorithm (GA). The experimental results show that a better prediction performance and generalization can be achieved with the proposed LSSVM compared to the traditional strategies such as SVM, BP neural networks, and logistic regression. As a result, we can conclude that the LSSVM is more suitable for the practical use in systemic financial risk predicting.