关系驱动的基于时间卷积网络的股票走势预测算法  

Relation-Driven Stock Trend Prediction Model Based on Time Convolutional Network

作　　者：赵戈 徐尚志[1] 李志鹏[1] 

机构地区：[1]同济大学电子与信息工程学院,上海

出　　处：《计算机科学与应用》2021年第10期2555-2567,共13页Computer Science and Application

摘　　要：由于股票市场的复杂性和波动性,股票价格走势预测被认为是一项具有挑战性的任务。基于图卷积网络(Graph Convolution Network, GCN)处理非欧式结构数据的能力,本文提出了一种股票走势预测算法。该算法通过时间卷积网络(Time Convolution Network, TCN)从股票的历史交易数据中提取特征,使用图结构对股票之间的关系进行量化,应用图卷积网络从相关股票中学习局部特征表示来预测股价走势。与以往的研究不同,本文强调个体特征和多尺度局部特征的融合,并应用小波包变换对数据进行去噪以提高算法的性能。本文提出的算法在中国大陆沪深300成分股数据集中进行了验证,在两个指标上取得了优于常见时间序列预测基线模型2%和3%的性能表现。Stock price movement prediction is considered a challenging task due to the complexity and volatility of the stock market. Based on the ability of graph convolution network (GCN) to process non-European structured data, this paper proposes a stock trend prediction algorithm. The algorithm extracts features from the historical trading data of stocks through TCN (Time Convolution Network), quantifies the relationship between stocks using graph structure, and uses GCN to learn local feature representation from related stocks to predict the movement of stock price. Different from previous studies, this paper emphasizes the fusion of individual features and multi-scale local features, and applies wavelet packet transform to denoise the data to improve the performance of the algorithm. The algorithm proposed in this paper is verified in the data set of CSI300 constituent stocks in the Chinese mainland. The performance of two benchmark indexes is 2% and 3% better than that of the common time series prediction baseline models.

关 键 词：图卷积网络 时间卷积网络 小波包变换 股票价格预测 

分 类 号：F83[经济管理—金融学]