融合CEEMDAN的SSA-BiGRU股价预测研究

Research on SSA-BiGRU Stock Price Prediction Integrating CEEMDAN

摘要: 由于股票市场的复杂性，股价预测精度一直不高。因此，本文结合信号处理领域的模态分解方法，建立了融合投资者情绪的多源异构信息股价预测模型。首先，该模型使用BERT获取财经新闻词向量，并使用改进后的融合自注意力机制的BiLSTM模型量化投资者情绪。其次，使用CEEMDAN算法分解股价序列为本征模态函数，最后，将投资者情绪、本征模态函数、历史交易数据、技术指标进行特征融合，通过融合自注意力机制的双向门控循环单元实现对次日股价的预测。本文所提出的预测模型在四支股票数据集上的拟合优度平均达到了97.39%，与现有的单一信息源预测模型相比，所提出的混合多种信息源的预测模型效果更加优越。

Abstract: Due to the complexity of the stock market, the accuracy of stock price prediction has always been low. Therefore, this paper combines the modal decomposition method in the field of signal processing to establish a multi-source heterogeneous information stock price prediction model that integrates investor emotions. Firstly, the model uses BERT to obtain financial news word vectors, and uses an improved BiLSTM model that integrates self-attention mechanism to quantify investor emotions. Secondly, the CEEMDAN algorithm is used to decompose the stock price sequence into intrinsic modal functions. Finally, investor emotions, intrinsic modal functions, historical trading data, and technical indicators are feature fused, and the prediction of the next day’s stock price is achieved through a bidirectional gated loop unit that integrates self-attention mechanism. The prediction model proposed in this paper has an average goodness of fit of 97.39% on four stock datasets, which is comparable to existing models. Compared with the single information source prediction model, the proposed prediction model that combines multiple information sources has superior performance.