Authors: Chih-Chieh Hung; Ying-Ju Chen; Siou Jhih Guo; Fu-Chun Hsu
Addresses: Department of Computer Science and Information Engineer, Tamkang University, Taiwan ' Department of Mathematics, University of Dayton, USA ' Department of Computer Science and Information Engineer, Tamkang University, Taiwan ' National Australian Bank, 700 Bourke Street, Melbourne, 3000, Vic, Australia
Abstract: Candlestick charts have been widely used to display price movements of a security, derivative, or currency for a specific period. They are one type of popular charts for day traders. Motivated by the conventional use of candlestick charts as a visual aid for decision making in stock, currency exchange, and commodity trading, we proposed a framework deep candlestick predictor (DCP) to forecast the price movements by reading the candlestick charts instead of reading the considerable body of numerical data from financial reports. DCP consists of three components: 1) chart decomposer: decomposes a given candlestick chart into several sub-charts; 2) CNN-autoencoder: derives the best representation of sub-charts; 3) 1D-CNN: forecasts the price movements. An extensive study is conducted by daily prices from Taiwan Exchange Capitalization Weighted Stock Index which contains 21,819 trading days. The result shows that DCP effectively achieves higher accuracy comparing to accuracy using conventional index-based models.
Keywords: candlestick chart; price movement prediction; convolutional autoencoder; convolutional neural network; CNN; deep learning.
International Journal of Ad Hoc and Ubiquitous Computing, 2020 Vol.34 No.2, pp.111 - 120
Received: 26 Mar 2019
Accepted: 28 Oct 2019
Published online: 17 Jun 2020 *