A Financial Data Analysis Method Based on Time-Series Generative Adversarial Network and Decomposition Learning

Stock price series are highly volatile and non-stationary, making it difficult to predict future movements. Deep learning technology has provided new ways to predict stock prices. In existing works, deep learning combined with stock price series decomposition is a common architecture. Inspired by this, we propose a financial data analysis method based on time series generative adversarial network (TimeGAN) and decomposition learning. Specifically, we use multi-view data, including the highest price, lowest price, opening price, closing price, and trading volume, as explanatory variables for predicting the closing price. In parallel, we utilize TimeGAN to generate data with similar distributions to these data to reduce overfitting due to data scarcity. Further, we decompose the closing price series into several subseries with relatively simple patterns by singular spectrum analysis (SSA) and then use self-attention mechanisms to let the subseries perceive each other and learn their dependency relationships, thus providing useful information about stock price change patterns for the prediction model. Experimental results on multiple national stock indices have demonstrated that the proposed method significantly improves prediction accuracy compared to recent advanced methods.