The Research Of Country Risk Assessment Based On Hybrid Neural Networks

With the rapid development of economic globalization, the internationalization of commercial banks in the world becomes more and more obvious. At the same time, the accompanying risk also becomes more and more diversified and country risk is the most concerned one. Based on the research on the definition of country risk, the comparative analysis of the qualitative and quantitative assessment methods, we set up the assessment index system of country risk and build the hybrid neural networks of country risk assessment in this article.There are two main parts in the empirical part of this article. One part is about the screening and processing of the indicators, the other part is about to use the neural networks to evaluate country risk and then to analyze the predictive ability of this method. On the aspect of the establishment of the index system, we select economic, political and social indicators on the basis of the definition of country risk and the country risk indicators of high frequency. Then we use the correlation analysis and logistic method to select variables. On the aspect of the establishment of the neural networks, we firstly use the two-class dependent variables as the output of the model to train the networks and built up the network model based on the principle of the multi-layer perception networks. The samples are divided into two grades:investment grade and speculative grade. The verification result of the holdout samples shows that the right classification percentage is100%. On this basis, we use the multi-class dependent variables as the output of the model to train the networks and built up the network model based on the principle of the probabilistic neural networks. The samples are divided into three grades. The percentage error of the training samples is0. The percentage error of the test samples is9.52%.In conclusion, the assessment result displays that EBI and UL have the best goodness of fit. The predicting outcomes of the multi-layer perception model verify that the predictive effect of the two-class networks is very good. But the right classification percentage of the three-class networks declines obviously.In general, the hybrid neural networks have better predictive ability because of the feature of functional approximation of this model, comparing with the traditional statistical models. But I don’t take country risk event of small probability into account in the empirical model of this article and don’t combine it with the international business. So this model needs to be improved further.