Abstract: In order to improve the predictive performance of Chinese stock indices, a hybrid model MVC5B (multi-channel-VMD-CBAM5-BiLSTM) is proposed, which integrates variational mode decomposition (VMD), convolutional block attention module (CBAM), and bidirectional long short-term memory network (BiLSTM). Unlike the commonly used decomposition-ensemble construction approach in hybrid models, MVC5B is constructed based on a proposed multi-channel input method. The multi-channel input method effectively avoids the cumulative errors and significant computational costs associated with the multiple predictions of decomposition-ensemble approach, thereby enhancing the predictive performance of MVC5B. Furthermore, the introduction of CBAM not only improves the predictive performance of the stock indices but also enriches the research on CBAM in stock index prediction. Empirical results based on multiple representative Chinese stock index datasets demonstrate that MVC5B exhibits significantly better predictive performance and simulated returns than popular forecasting models. Additionally, the empirical results further confirm the superiority of the multi-channel input method over the decomposition-ensemble approach and the effectiveness of CBAM in stock index prediction.

Key words: stock index prediction, convolutional block attention module (CBAM), bidirectional long short-term memory network (BiLSTM), variational mode decomposition (VMD), multi-channel input method

摘要： 为了提高中国股指的预测表现，提出一个融合了变分模态分解（VMD）、卷积注意力模块（CBAM）和双向长短期记忆网络（BiLSTM）的混合模型MVC5B（multi-channel-VMD-CBAM5-BiLSTM）。不同于混合模型常用的分解-集成构造方法，MVC5B基于提出的多通道输入方法构造而成。多通道输入方法基于自身一次性预测的特点可以有效规避分解-集成方法多次预测带来的累计误差和巨大计算成本，从而提升MVC5B的预测表现。CBAM的引入不但提升了股指的预测表现，而且还丰富了股指预测问题中关于CBAM的研究。基于多个具有代表性的中国股指数据集的实证结果显示MVC5B的预测表现和模拟收益显著优于流行的预测模型。实证结果还进一步证实了多通道输入方法相比于分解-集成方法的优越性以及CBAM在股指预测问题中的有效性。

关键词: 股指预测, 卷积注意力模块, 双向长短期记忆网络, 变分模态分解, 多通道输入