基于TBA融合模型的股票指数预测  

Stock index forecasting based on TBA fusion model

作　　者：韩迪 郭维 廖凯 孙传一 汪勃澄 林坤玲 HAN Di;GUO Wei;LIAO Kai;SUN Chuanyi;WANG Bocheng;LIN Kunling(School of Gredit Management,Guangdong University of Finance,Guangzhou 510521,Guangdong Province,P.R.China;Finance Credit Big Data Research Center,Guangdong University of Finance,Guangzhou 510521,Guangdong Province,P.R.China)

机构地区：[1]广东金融学院信用管理学院,广东广州510521 [2]广东金融学院金融信用大数据研究中心,广东广州510521

出　　处：《深圳大学学报（理工版）》2023年第6期665-673,共9页Journal of Shenzhen University(Science and Engineering)

基　　金：广东省自然科学重点领域专项基金资助项目(2020ZDZX3066);广东省哲学社会科学规划资助项目(GD21YYJ02,GD19CLJ01)。

摘　　要：由于股票市场的波动性和复杂性特点,股指预测一直是金融预测研究中的难点.长短期记忆(long short-term memory, LSTM)网络模型常用于金融指数的预测中,但该模型在长时间序列上易导致数据信息利用不充分.利用双向长短期记忆(bidirectional LSTM, BiLSTM)网络模型、时间卷积网络(temporal convolutional network, TCN)和注意力机制协同提高了模型识别以及提炼长时间序列数据特征的能力,构建一种新型股指预测融合模型TCN-BiLSTM-attention(简称TBA模型).以中国境内近30年的公开股指数据集为例,将TBA模型与目前金融类主流的机器学习、神经网络预测算法以及kaggle竞赛排行前列的模型在上证指数、沪深300指数与创业板指数进行预测对比和消融实验.结果显示,相较于对照实验组的平均预测误差,TBA模型有明显降低且表现稳定,兼具准确性与鲁棒性.研究结果可广泛用于基于时间序列的多种金融预测场景.Due to the volatility and complexity of stock market,stock index prediction has always been a challenge in the field of financial forecasting.Long short-term memory(LSTM)network model is commonly used in financial index forecasting,however,this model has some limitations in long time series which may lead to insufficient use of data information.By using bidirectional long short-term memory network model(BiLSTM),temporal convolutional network(TCN)and attention mechanism,a novel fusion model named TCN-BiLSTM-attention(hereinafter referred to as TBA model)for stock index forecasting is constructed to further improve the ability of recognition and extraction of long time series data features of the model.Taking the public stock index datasets within China for nearly 30 years as an example,the TBA model is compared and ablated with the current mainstream machine learning and neural network prediction algorithms in finance as well as the top ranked models in Kaggle.The experimental results show the TBA model has significantly lower average prediction error and more stable performance compared with the average error baseline of the control experimental group in the forecasting of SSE,CSI 300 and GEI multi-day indices,thus this model can be used in a variety of financial forecasting scenarios based on time series.

关 键 词：数字经济 股指预测 长短期记忆网络 时间卷积网络 注意力机制 消融实验 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] F208[自动化与计算机技术—控制科学与工程] TP3391[经济管理—国民经济]