Bitcoin Price Fluctuation And Forecasting Methods Research Based On Empirical Mode Decomposition

Bitcoin is a highly controversial and disruptive digital currency.It is a wonderful combination of computer technology,cryptographic principles,and liberal economic ideas.It has features such as a limited amount,decentralization,open and transparent transaction records,and other credit currencies that are not available.Bitcoin has commodity attributes and currency attributes that can be used for investment speculation and for payment transactions.The fluctuation of Bitcoin prices are also differs from the price changes of traditional currencies and commodities.Therefore,analyzing and predicting the trend of Bitcoin prices has become a focus of attention in the theoretical and investment communities.Since 2013,Bitcoin prices have experienced several major declines due to a variety of factors.With the rise of trade protectionism and the slowing down of world economic growth,bitcoin investment has gradually become one of the tools for people's financial management and investment.It is extremely important to analyze and determine the price and trend of Bitcoin.Bitcoin investors only have a deep understanding of the mechanism of Bitcoin price fluctuations and Bitcoin price movements can effectively circumvent the risk of fluctuations in Bitcoin prices,and then get the ideal investment income.In this paper,a bitcoin price analysis and prediction method based on integrated empirical mode decomposition(EEMD)is proposed.In this paper,bitcoin price series from January 4,2013 to April 12,2017 and the daily closing price data as the research object,the first bitcoin price descriptive statistical analysis to understand the basic characteristics of bitcoin prices,followed by the use of EEMD technology will bitcoin price series is decomposed into 10 different frequency components(including nine intrinsic mode(IMF)and an error term).First,this paper analyzes the decomposed components as follows: 1,Calculate the average period of each IMF 2,Analyze the correlation between IMFs and bitcoin price original sequences and their contribution to variance.Secondly,the reconstruction algorithm put these different frequencies and sizes of IMFs stack into three new components,different frequencies can be divided into: low frequency component and high frequency component and trend component,this paper from the three time scales are bitcoin price time series theory analysis and analysis of characteristics.The conclusion of this paper is that the price of bitcoin is mainly influenced by long-term trends,major events and short-term market fluctuations.Because the price of bitcoin is a nonlinear and nonstationary time series,it is a challenging job to accurately predict the trend of its price.The general statistics and econometrics models are based on the hypothesis that the data is linear.It is extremely difficult to extract the nonlinear mode hidden in bitcoin price.Therefore,we can not get the exact prediction result of bitcoin price.In order to break through the defect of statistics and econometrics model,this paper tries to propose a EMD-LSTM model prediction method based on EEMD and LSTM(long and short term recurrent neural network).The method uses EEMD to bitcoin price series is decomposed into several different frequency and scale components,each component according to the frequency of different form three new sequences,they represent short-term market volatility,mid major events,long-term trend;the three sequence,construction parameters of LSTM model respectively the predicted three predictive value sequence.The root mean squared error(RMSE),the mean absolute percent error(MAPE)and the average absolute deviation of MAD(Mean Absolute Deviation).The prediction error size of each model in the experiment is measured.Finally,the final prediction value is added to the three predictive value sequences.In this paper,the price data of bitcoin is used to verify the effectiveness of this method and to predict the price of special currency.The results show that the prediction results used in this paper are better than the single LSTM model.