Forecasting Financial Time Series Based On ICEEMDAN-GWO-MKELM

As an important scientific research object,time series plays an important role in all corners of social production.For example,the time series of crude oil prices is related to various industrial production and social activities that rely mainly on its processed products.In the time series,financial time series is the focus of attention of all parties including investors,government managers and scholars.However,financial time series is often affected by various factors,such as macroeconomics and policies,geopolitics,and emergent information and with characteristics of nonlinearity,non-stationarity,and multi-scaling.Thus,financial time series is extremely uncertain and difficult to predict.How to construct an effective and accurate financial time series forecasting model has great theoretical and practical value for related investors,researchers,enterprises and governments.The main research content and results of this thesis are as follows: Use ICEEMDAN(Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise),GWO(Gray Wolf Optimization),MKELM(multi-Kernel Extreme Learning Machine,multi-core extreme learning machine)theoretically to construct ICEEMDAN-GWO-MKELM ensemble prediction model,in which the multi-kernel extreme learning machine is proposed based on KELM(Kernel Extreme Learning Machine).The research data selected in this thesis are derived from the international crude oil time series of WTI and BRENT respectively,and have come from the exchange rate time series of US dollars,Euros,and 100 Yen against RMB in China Foreign Exchange Trading Center.The model first uses ICEEMDAN to decompose the selected financial time series,then uses the GWO optimized MKELM to predict the decomposed sub-sequences respectively,and finally sums the predicted values of the sub-sequences to obtain the overall predicted value.The effectiveness of this model is verified by comparison with other advanced prediction models and Nemenyi test on the prediction results.On this basis,the thesis also studied how the prediction performance of the proposed ICEEMDANGWOMKELM model is affected by some model parameter settings through the controlled variable method,and on this basis proved the reliability of our model.The main contributions of this thesis are: 1.A new time series forecasting method is proposed,which shows better forecasting accuracy on financial time series including crude oil prices and exchange rates,and can be considered to be used in the fields of production,research,and government management as a time series prediction model with strong practicability;2.On the basis of kernel extreme learning machines,an idea of introducing multiple kernels to enhance the model’s predictive ability is proposed,and a natural calculation method is used,which is gray wolf algorithm,to optimize the parameters of the multi-kernel extreme learning machine,so as to realize the optimization of the multi-kernel extreme learning machine;3.It is proved that the prediction of financial time series based on the "divide and conquer" idea is effective.By decomposing the complex original time series,sub-sequences which are simple and retain the key information of different dimensions of the original time series can be obtained.The relatively simple subsequence helps reduce the difficulty of forecasting.Retaining the characteristics of the different dimensions of the original time series also ensures that the model will not miss the key information for the forecasting effect.Therefore,it can finally help the prediction model to obtain excellent prediction accuracy.