Open AccessStock Market Prediction Based on Time-frequency Analysis and Convolutional Neural Network
Author(s) -
Dandi Jia,
Qiang Gao,
Hui Deng
Publication year - 2022
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2224/1/012017
Subject(s) - time–frequency analysis , spectrogram , short time fourier transform , convolutional neural network , stock market , computer science , pattern recognition (psychology) , fourier transform , artificial neural network , classifier (uml) , stock market index , stock (firearms) , econometrics , artificial intelligence , algorithm , mathematics , fourier analysis , materials science , telecommunications , radar , mathematical analysis , paleontology , horse , metallurgy , biology
Recently, researchers have shown an increased interest in stock market prediction with neural networks. Stock market is affected by a multiplicity of factors with different active periods, thus financial time series possess multiscale frequency characteristics, which can be exploited to facilitate prediction of stock market. In this paper, we propose a stock market prediction model combining time-frequency analysis and convolutional neural network (CNN), in which the influence extent of different frequency components has been considered. We transform original financial time series into the spectrogram reflecting time-localized frequency information by short-time Fourier transform (STFT). The 2-dimensional time-frequency feature is obtained from the spectrogram by frequency bands extraction, which is then pre-weighted and input into CNN to forecast the future price change. The frequency bands extraction and pre-weight are set according to the frequency influence. The results of experiments on Shanghai Composite Index show that the proposed model with frequency bands extraction considering frequency influence achieves a 4% relative decrease in mean absolute error (MAE) compared with that does not consider the frequency influence. Moreover, the pre-weight gives an additional 3% relative decrease of MAE.