基于Transformer模型的多尺度股票预测

Multi-Scale Stock Prediction Based on

随着经济全球化所带来的股票交易市场扩大,股票交易数据日渐激增,如何使用更有效的方法在众多股票数据中选取优质股成为股民愈发关注的问题。而采用人工智能技术来分析和处理交易数据可以有效地辅助股民进行合理的股票选择。近些年,不同学者尝试使用人工智能技术的深度学习方法来解决股票预测问题,虽然取得了不错的效果,但是缺少对不同预测周期的股票价格预测能力的稳定性。本文为解决以上存在的问题,采用深度学习技术中的变换神经网络(transformer)作为网络基本架构对股票价格的开盘价进行预测,将不同时间尺度的股票价格输入到网络模型中,能够处理更全局的时间序列特征信息,然后通过CNN技术将不同尺度的特征进行融合,使得模型能够满足不同周期数据的预测变化。本研究选取乐普医疗(300003)、亿纬锂能(300014)、浦发银行(600000)和东风汽车(600006)四组股票作为实验数据。实验结果表明,在5个基本交易指标(开盘价、最高价、最低价、收盘价、成交量)的影响因素下,本研究的多尺度股票价格预测模型具有更好的性能及泛化能力。

With the expansion of the stock trading market brought about by economic globalization, the stock trading data is increasing day by day. How to use a more effective method to select high-quality stocks from a large number of stock data has become a more and more concerning issue for inves-tors. The use of artificial intelligence technology to analyze and process transaction data can effec-tively assist investors in making reasonable stock selections. In recent years, different scholars have tried to use the deep learning method of artificial intelligence technology to solve the problem of stock prediction. Although they have achieved good results, they lack the stability of stock price forecasting ability in different forecasting periods. In order to solve the above problems, this paper uses the transform neural network (transformer) in the deep learning technology as the basic net-work structure to predict the opening price of the stock price, and input the stock price of different time scales into the network model, which can deal with more global. Then, the features of different scales are fused by CNN technology, so that the model can meet the prediction changes of different periodic data. This study selects four groups of stocks of Lepu Medical (300003), Yiwei Lithium En-ergy (300014), Shanghai Pudong Development Bank (600000) and Dongfeng Motor (600006) as experimental data. The experimental results show that under the influence of five basic trading in-dicators (opening price, high price, low price, closing price, and trading volume), the multi-scale stock price prediction model of this study has better performance and generalization ability.