Singular Spectrum Analysis based Long Short-Term Memory for Predicting Bitcoin Price

Abstract

Bitcoin, a leading cryptocurrency in the financial market, is full of non-linearity, non-stationarity and high volatility. To make risk management strategies, emphasis on cryptocurrency price predicting is truly needy. However, studies about cryptocurrency prediction are lacking. In this paper, a novel hybrid model combining long short-term memory (LSTM), a state-of-the-art sequence learning method, with singular spectrum analysis (SSA) was proposed to predict Bitcoin price. SSA was employed to decompose the original time series into independent signals in term of trend, market fluctuation and noise. A smoothed series with valid information was reconstructed with reduction of noise. By introducing the smoothed series sequence into LSTM, prediction value is obtained. Empirical analysis shows that the proposed hybrid SSA-LSTM model outperforms baseline single LSTM model, according to root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). The result suggests that the proposed hybrid model has satisfactory ability to grasp pattern of Bitcoin price series since SSA can extract valid information from the original series and avoid overfitting.