Centralized decomposition approach in LSTM for Bitcoin price prediction

Abstract

It has been reported that integrating time-series decomposition methods and neural network models improves financial time-series prediction performance. Despite its practical importance, the prediction performance of cryptocurrency prices, including Bitcoin, at the tail domain of the label distribution is generally less successful than the mean performance across the entire domain of the label distribution. In order to enhance the overall predictive performance of the Bitcoin price, we propose the Centralized Clusters Distribution (CCD) as a novel input data filtering mechanism that significantly improves both the tail performance and the overall performance by mitigating the extreme bimodality inherent in Bitcoin price. The combination of CCD and the Weighted Empirical Stretching (WES) loss function, which imposes different penalties depending on the label distribution, outcomes in an additional performance gain. In the Long-Short Term Memory (LSTM) and the Singular Spectrum Analysis (SSA) decomposition method, the CCD-WES strategy outperforms the native experiment by 11.5% and 22.5% Root Mean Square Error (RMSE) gain in the whole and extreme domains of the label, respectively.