The Comparison of Some Algorithm Based on Ceemdan

Stock price prediction is an indispensable part of the investment market and lots of approaches have been proposed. This article implements Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose the dataset, because the time series data were usually nonlinear and nonstationary. Meanwhile, our algorithms do not have to decide the basic function. In addition, different algorithms are adopted to process different levels of Intrinsic Mode Function (IMF) in order to capture their different characteristics. In that case, a combined algorithm containing CEEMDAN, MLP-BP, LSTM/GRU and Linear Regression is used. The combined algorithm works well on Tesla’s price series from 2010 to 2020, and it can not only process the data with a high accuracy, but also has a short running time because some of the algorithms in the combination are very simple. Then the combined algorithm is adopted on a more representative dataset called S&P 500 index. It turns out it is still very accurate and fast.