多时间尺度下变体生成式对抗网络的股价预测

Stock price prediction with a variant generative adversarial network in multiple time scales

股价预测能为公司经营、投资决策和市场监管提供重要依据。【目的】为了避免特征提取不足与预测不准等问题,我们构建了多时间尺度下变体生成式对抗网络对股价涨跌方向进行预测。【方法】首先以双向长短期记忆网络构造生成器,以卷积神经网络构造判别器;然后分别对生成器与判别器在多时间尺度数据上进行博弈训练,提取长期与短期特征后将结果拼接;最后获得预测模型。【结果】选取沪深300指数、建设银行与陕西煤业股价为样本进行实证分析,试验发现沪深300指数涨跌预测准确率达到59.63%,个股数据验证表明本文模型具有一定的稳定性与优越性。【结论】本模型能提高预测股价涨跌的准确率,丰富了金融数据分析方法。

The stock price prediction can provide an important basis for the company’s operation, investment decision-making and market supervision. [Objective] To avoid the problems of insufficient feature extraction and inaccurate prediction, a variant generative adversarial network in multiple time scales was constructed to predict the rises and falls of stock prices. [Method] First, the generator was constructed with the bidirectional long and short-term memory network, and the discriminator was constructed with the convolutional neural network;then, game training was conducted on multi-time scale data by the generator and the discriminator respectively;finally, the prediction model was obtained by splicing the results after long-term and short-term features were extracted. [Result] The CSI 300 Index, China Construction Bank and Shaanxi Coal Industry stock prices were selected as samples for empirical analysis. It is found that the accuracy rate of CSI 300 Index rise and fall prediction reaches 59.63%. The validation of individual stock data shows that the proposed model boasts certain stability and superiority. [Conclusion] This model can improve the accuracy of predicting the rises and falls of stock prices, and enrich the financial data analysis methods.