Research On RMB Exchange Rate Forecasting Based On Comprehensive Ensemble Learning

In the current complex and volatile background of the global political landscape and economic environment,the internationalization of the RMB is facing numerous challenges.The external investment environment is turbulent,such as trade protectionism,populism,and the Federal Reserve's interest rate cuts.In 2020,the global epidemic spreads and the economic environment has been repeatedly impacted,which has led to violent fluctuations in the RMB exchange rate price.At the same time,the time series of the RMB exchange rate presents obvious characteristics of nonlinearity,non-stationary,high complexity and high noise.Therefore,a reasonable and accurate analysis of the RMB exchange rate trend is related to the investment decisions of investors,enterprises and individuals to a certain extent.According to the existing research on RMB exchange rate forecasting,the forecasting effect based on a single time series model or artificial intelligence model is often unable to fully dig out the characteristics and laws of the RMB exchange rate series,but the comprehensive ensemble strategy is a popular forecasting method in existing research.Based on the analysis of the advantages and disadvantages of the existing exchange rate forecasting research,this paper proposes two comprehensive ensemble forecasting frameworks.The specific contents are as follows:(1)RMB exchange rate prediction framework based on multi-model combination ensemble.The framework mainly considers that various forecasting models have their own advantages,and finally three representative forecasting models are selected.Specifically,linear feature prediction accurate time series model(ARIMA),nonlinear fitting better neural network model(KELM)and complex dynamic feature mining better deep learning model(Bi GRU).At the same time,considering that useful information may remain in the prediction errors,error correction method is adopted to forecast the prediction error of the three models.Among them,the hyperparameters of the KELM and Bi GRU models are dynamically and adaptively selected through the tree-structured Parzen estimation algorithm(TPE).Final,the prediction result of the RMB exchange rate is obtained through the averaging method.(2)The RMB exchange rate prediction framework based on the decomposition reconstruction ensemble method.The framework mainly includes four parts: exchange rate sequence decomposition,component feature analysis,component reconstruction and forecast ensemble.Four decomposition algorithms of EMD,EEMD,CEEMDAN and ICEEMDAN are used in the decomposition of exchange rate series.In the component feature analysis,the complexity,the correlation with the original sequence and the average period of each decomposition component are analyzed.In the component reconstruction,based on the analysis results of the characteristics of each component,this paper uses the fuzzy C-means algorithm to combine the components with high similarity,and reconstruct them into three interpretable components.In the prediction ensemble,this paper selects an appropriate prediction model for each reconstruction component and uses error correction method for the first two components with higher complexity to improve the prediction accuracy.The final prediction result is obtained by linear ensemble of the prediction values of each component.In the empirical analysis,this paper applies two comprehensive ensemble frameworks to the forecast of the US dollar against the RMB central parity and offshore closing price.And by predicting the daily maximum and minimum offshore prices to obtain the fluctuation range of the daily closing price.In order to verify the prediction accuracy and generalization ability of the proposed framework,a multi-input and multioutput strategy is adopted to carry out multi-step forward prediction.According to the results of error evaluation,statistical analysis and interval evaluation,the two forecasting frameworks that proposed in this paper can achieve better forecasting performance.In contrast to the two,the prediction results obtained by the decomposition reconstruction ensemble method are more stable,which can provide a more reasonable reference basis for policy makers and investors.