| With the continuous development of China’s stock market and the gradual improvement of financial systems,more and more people are joining the ranks of stock investment.Over the past three decades,China’s stock market has accumulated a large amount of trading data,and stock movement prediction has received increasing attention from investors and researchers.As a classic problem in the field of computer science and finance,correct stock movement prediction not only reduces risk and improves investment returns for investors,but also has a positive impact on the healthy and stable development of the overall securities market.Due to the complexity of the stock market,its price series exhibit highly volatile and nonlinear characteristics.Most of the current stock movement prediction methods directly use interday data as input,and the prediction models use supervised deep learning models.However this leads to two problems.Firstly,stock price data is highly stochastic and inter-day data cannot truly reflect the changes of stock prices on that day,however,when the sampling frequency increases,the length of data not only becomes longer but also brings in a lot of noise.Secondly,due to the weak generalization of the model,existing supervised learning-based stock movement prediction methods are prone to overfitting and the model fails to learn effective representations.In this thesis,we propose a contrastive learning based stock movement prediction method CLSR(The Framework for Contrastive Learning of Stock Representations),which deals with both intra-day and inter-day series separately.For intra-day minute-level data,two augmented views are obtained by random data augmentation.These two augmented views are subjected to a newly constructed hybrid encoder composed of Transformer and Temporal Convolution Network to extract the global and local features in the intra-day series.Then,they are trained by a contrastive learning network to obtain valid intra-day features.For the interday data,CLSR uses Long Short-Term Memory network with an attention mechanism to obtain long-term trend information from the intraday data.Finally,the inter-day features are fused with the intra-day minute-level features to obtain the final stock representation.In addition,to fully validate the effectiveness of contrastive learning in stock movement prediction,we propose the LCPC model.LCPC divides the stock price series into several sub-series,and then learns the contextual relationships between subseries by contrastive predictive coding to model the state transitions between stock price subseries.We back-tested historical data on the CSI 500 and DJIA components and experimentally proved that the proposed method outperforms the latest stock movement prediction schemes.The main contributions of this thesis:1)The CLSR model is proposed,which introduces a contrastive learning method based on the traditional supervised learning,and solves the weak generalization problem of the model by mining the similarity between samples.2)The hybrid coding network designed for single-day long series stock price data can extract intra-day information from both global and local aspects to better capture the intra-day market changes.3)Obtaining long-term historical information of stock prices through the historical state network and mining the periodicity and trend of daily frequency data.4)Designing the LCPC model to model the state transitions between stock price subseries,we verify that the CLSR network is more suitable for dealing with the Chinese stock market with high complexity and the LCPC network is more suitable for the US stock market with relatively complete market,which provides a new idea for stock movement prediction research. |
