Research On Bitcoin Price Time Series Based On Machine Learning Technology

In 2008,Satoshi Nakamoto proposed the concept of bitcoin.After a period of development,the open source software that came into being according to its concept,and research on the bitcoin architecture and its core algorithms formed the bitcoin ecological environment.Due to its early features of easy circulation,low transaction costs,easy mining,and decentralization,Bitcoin developed rapidly in the early days.However,due to the anonymity of the bitcoin system,this was used by criminals for financial criminal activities.Governments of various countries once strictly controlled bitcoin transactions,resulting in the instability of bitcoin prices.After years of ups and downs,bitcoin has gradually been accepted by the majority of investors,the bitcoin market has gradually stabilized,and the impact of external policy factors on the bitcoin market has gradually become smaller.Therefore,it is of practical significance to study the influence of economic factors and internal factors of Bitcoin on the formation of Bitcoin price.Bitcoin has considerable research value in the fields of economics,cryptography,and computer science due to its inherent characteristics of combining encryption technology with monetary units.In recent years,people have conducted a lot of modeling studies on the time series of Bitcoin prices as a new market variable with specific technical rules.In the existing research,it is mainly studying the impact of several factors such as market power,investor attractiveness,and global macro financial factors on the formation of Bitcoin price.However,the core technology of Bitcoin is the formation of the blockchain,which is the key point that Bitcoin is different from other financial products.It is directly related to the supply and demand relationship of Bitcoin.According to existing research,in addition to macroeconomic factors,blockchain information and blockchain network information are rarely systematically used to study and describe the process of bitcoin price formation.In view of the above problems,this paper intends to use the blockchain information,blockchain network information,global currency market indicators,and macroeconomic factors to model and predict the Bitcoin price series,and systematically evaluate and characterize the Bitcoin price series.Based on the existing research,this paper collects data in 26 dimensions including the Bitcoin network block propagation delay,hash rate,Bitcoin mining difficulty,the US Dow Jones index,and the US dollar-euro exchange rate.Covers macroeconomic factors,global money market indicators,blockchain information,and blockchain network information.This article first uses a linear regression model to perform a linear analysis on the formation of bitcoin price.After experimentally analyzing and excluding feature attributes with multiple colinearity features,the model is modeled using a ridge regression model and used as a control group for the experiment to analyze various features.The linear correlation between vector and bitcoin price.Residual analysis was performed on the prediction results,which proved the validity of the linear regression model.Then model the long and short-term memory neural network model for the variable set after excluding multicollinearity and the variable set without excluding multicollinearity,and choose different time steps for analysis.Twenty data sets were formed by combining step size and variable set,and comparative experiments were performed respectively.After the experiment,it is found that the modeling effect is best when a data set formed by a time step of 4 and a multicollinearity variable set is not excluded.The experiment proves that the future price formation of Bitcoin can be estimated and analyzed by the model by learning historical data,and the data selected four days before the research time point for analysis is the best.At the same time,through experimental comparison,selecting a data set that does not exclude multicollinearity for modeling training is more effective than selecting a data set that excludes multicollinearity,and the experimental effect is significantly improved.This also proves the long-term and short-term memory neural network from the side.The model can effectively learn the non-linear information in the data,and the nonlinear fitting ability of the model is relatively strong.Finally,based on the results of the long-term and short-term memory neural network model experiments,a Bayesian neural network model was selected using a data set consisting of a time step of 1 and a set of variables that did not exclude multicollinearity.Experiments have found that the Bayesian neural network model has a good interpretability for the Bitcoin price series,and it has a good predictive performance for the Bitcoin price and price volatility.Bayesian neural networks can evaluate the uncertainty of estimates and predictions by analyzing the posterior distribution of the parameters.In the end of this paper,the mean value of the posterior sampling of Bayesian parameters is selected,and the degree of influence of neural network input variables on the output of the neural network is studied and calculated using the variable sensitivity analysis method based on the connection weight of neural network proposed by Professor Garson [1].The experimental results show that blockchain information is the most critical factor affecting the formation of Bitcoin price,followed by macroeconomic indicators and global money market indicators.In terms of blockchain information,the total transaction cost of Bitcoin is the factor that has the greatest impact on the price of Bitcoin,followed by the transaction volume of Bitcoin.In addition,factors such as the exchange rate of the US dollar to the euro,the hash rate,the difficulty of mining,and the delay of block propagation all have varying degrees of influence on the formation of bitcoin prices.