面向图片和文本多源异质数据的股票预测联合模型

A Joint Model for Stock Prediction Based on Image and Text Multisource Heterogeneous Data

摘要: 股票预测一直是金融界研究的热点问题，近年来融合文本、图片这类非结构化数据成为提高预测精度的研究方向。本文建立了一种能够同时处理多源异质数据的股票价格走势预测联合模型，分析词云图片、股吧评论文本和股票交易数据。联合模型分为两个分支，一个分支运用CNN模型分析由股民评论文本转为的词云图片，另一分支运用LSTM模型处理历史股票交易数据和由股民评论文本得到的情感评分，两个分支共同预测4天、6天、8天的股票走势涨跌。结果表明使用词云图片的CNN模型表现优于情感分析的LSTM模型，证明词云图片的可使用性，且联合模型结果优于两个单一模型，准确率稳定在0.6~0.7之间。

Abstract: Stock prediction has always been a hot topic in the financial industry, and in recent years, integrat-ing unstructured data such as text and images has become a research direction to improve predic-tion accuracy. This article establishes a joint model for predicting stock price trends that can sim-ultaneously process multi-source heterogeneous data, including the analysis of word cloud images, stock bar comments, and historical stock trading data. The proposed model comprises two branches. The first branch utilizes a CNN model to dissect word cloud images derived from stock comment texts, while the second branch employs an LSTM model to process historical stock trading data and emotional scores gleaned from the stock comment texts. Two branches jointly predict the rise and fall of stock trends for 4, 6, and 8 days. The findings indicate that the CNN model’s use of word cloud images yields superior performance compared to the LSTM model’s sentiment analysis. This out-come underscores the efficacy of leveraging word cloud images as a predictive tool. Moreover, the joint model’s results surpass those of the individual models, with an achieved accuracy consistently ranging between 0.6 and 0.7.