The Early Warning System Under Boosting Algorithm

Due to the devastating and contagious nature of the financial crisis,the Early Warning System(EWS)has received constant attention from both policymakers and researchers.Although it is difficult to predict the currency crisis accurately,there is no doubt that developing the Early Warning System is the top priority of financial risk control.It will prevent the country from falling into financial crisis.This paper uses the traditional EWS,including logit regression model,multinominal logit regression model,decision tree,etc.,as a basic model.Then it applies a boosting algorithm,including AdaBoost,SAMME and SAMME.R,to develop a new Early Warning System to predict currency crisis.From the results of the traditional model,the accuracy and stability of the crisis prediction have some limitations.Therefore,this paper considers using a boosting algorithm to improve the model's prediction ability.This paper selects an effective boosting algorithm,AdaBoost,and then selects the SAMME and SAMME.R algorithm for multiple classification situations.The superposition of AdaBoost's basic model will make the training error decrease at an exponential rate,and it is often effective in avoiding overfitting in practical problems,which will increase the stability of the prediction result.This paper selects data from 15 emerging market countries including Asia,Africa and Latin America from January 2006 to October 2017.The empirical results reveal that the prediction results of the boosting algorithm show significant performance when there is a crisis.Among them,including the basic model,the SAMME algorithm,the SAMME.R algorithm,and the multiple models with adjusted sample weights,the multiple decision tree(three-tier)model under SAMME algorithm has a substantial improvement in ability to forecast currency crisis.Therefore,this paper supports that the prediction result represented by AdaBoost on the EWS problem is superior to the traditional model to some extent.Finally,the sample countries are analyzed by applying the best performing boosting model.The empirical results show that the overall accuracy of the crisis is correctly pre-warned by 90%(45/50)throughout the sample interval.In terms of the study period,except for Russia,China,and Argentina,every crisis in other countries can be accurately predicted.In comparison,the crisis prediction of China is non-prominent(7 times conflicts with 5 times predictions).The possible reason may be that although the control of exchange rate is increasingly liberalized,the prudence in the exchange rate policy remains unabated in the process.It is still more special than other countries.In general,the multiple decision tree(three-tier)model under SAMME algorithm has improved widely applicability to sample countries.It has proved few restrictions on country selection.