Bitcoin Price Prediction Based On News Headline Text Mining And Deep Learning

As one of the most popular cryptocurrencies,Bitcoin has attracted a great deal of attention from most people in the investment market,and investors have gradually increased their investment in Bitcoin.However,the characteristics of Bitcoin's decentralization and anonymity of transactions also make it very convenient for some illegal groups to conduct illegal money laundering and other illegal activities.Therefore,in order to guarantee the state's supervision of Bitcoin to a greater extent and ensure the safety of investors' funds,it is very necessary to predict the price of Bitcoin.In this thesis,the daily closing price of Bitcoin is taken as the research object for analysis and prediction,and the Bitcoin trading volume,a variety of major exchange rates,a variety of indexes,blockchain technical indicators,Bitcoin-related news,Google trends and other indicators from January 4,2016 to December 31,2018 are selected to construct a characteristic system.It needs to use sentiment analysis,LDA topic model and other text mining methods to numerical news text.In terms of prediction model,this thesis constructed a new hybrid prediction model,LSTM-AM-CNN,which integrates long short term memory,convolutional neural network and attention machine.Firstly,LSTM is used to extract time features from input data.Then,the attention mechanism allocates different attention weights to the output sequence of LSTM according to the importance degree.Finally,local features are further captured through CNN.In order to verify the better prediction performance of LSTM-AM-CNN model,three benchmark models LSTM,LSTM-AM and LSTM-CNN were selected in this thesis.The results show that the performance of the proposed model on RMSE,MAE and MAPE is better than that of the benchmark model.In addition,the effectiveness of introducing emotional features in improving the accuracy of Bitcoin price prediction was confirmed by comparing the prediction results of the above model with two groups of different feature combinations(with and without emotional features).