An Empirical Study On The Prediction Of Cryptocurrency Price Trend By Time-series Deep Learning Model

With the continuous development of blockchain technology and the rapid rise of bitcoin price,cryptocurrencies have begun to enter people's vision.There are thousands of cryptocurrencies on the market now.The special attributes of cryptocurrencies make it suitable for both investment and speculation,as well as for payment,and their price fluctuations are also different from those of traditional currencies and financial products.Therefore,predicting the price trend of cryptocurrency has become a hot topic of research.Traditionally,investors usually make price predictions based on statistical analysis or simple machine learning.Due to the non-linear,high-noise nature of cryptocurrency prices,these methods have limitations and not very effective.In recent years,artificial neural network has a good performance in natural language processing,automatic driving,computer vision and so on because of its strong ability of studying and non-linear modeling.Now more and more people are considering using artificial neural network to predict price trends.Among them,long short-term memory network(LSTM),as a member of recurrent neural networks,has attracted everyone's attention with its powerful sequence modeling capabilities.Temporal convolutional network(TCN)is a new type of convolutional neural network,which has been proposed recently.It defeats the recurrent neural network in a number of sequential tasks by virtue of its special structure.In addition,as an unsupervised learning model,stacked denoising autoencoder has a strong advantage in feature extraction.Based on the feature extraction of stacked denoising autoencoder and the excellent time-series data processing capability of temporal convolutional network,this paper makes method innovations in cryptocurrency price prediction.The main research work of this paper is as follows:1.Considering the relevant factors of cryptocurrency price,this paper selected15 characteristics of opening price,closing price,high price,low price,trading volume and other technical indicators,and established an index system for prediction of cryptocurrency price.2.In the unsupervised feature learning stage,in order to learn the deep features of the cryptocurrency price,this paper establishes an unsupervised learning model,stacked denoising autoencoder.3.In the supervised prediction stage,the paper introduces the temporal convolutional network and adjusts the parameters of the temporal convolutional network to improve the prediction effect.Finally,the temporal convolutional network based on stacked denoising autoencoder(SDAE-TCN)was proposed,and the price data of Bitcoin and Ethereum were selected as research objects.The LSTM,GRU,TCN,SDAE-LSTM are constructed and compared with the model proposed in this paper.The empirical results show that the stacked denoising autoencoder can effectively improve the accuracy of the model,and the SDAE-TCN proposed in this paper can achieve the best prediction effect.4.In order to improve the prediction ability of the model,this paper combines the proposed model with the attention mechanism.The empirical results show that the attention mechanism can effectively improve the prediction accuracy of the model by assigning weights to the data output by the TCN.