统计与数据科学学院知识库

近年来，我国股票证券市场开始蓬勃发展，套利策略逐渐在量化交易的发展 中实现从理论研究到制度设计再到实践探索的基本过程，是我国证券市场发展的 重要里程碑。随着金融市场交易数据的指数爆炸式增长，影响股票价格的因素也 越来越多，因此，对分析金融市场技术工具的要求也越来越高。计算机技术的发 展使得机器学习方法进入大众视野，深度学习技术是机器学习中的重要组成部分， 在处理非线性关系和高维数据上都有着巨大的优势。 长短期记忆网络模型（LSTM）因其出色的递推性质，适合处理时间序列相 关的问题。但 LSTM 模型具有其独特的结构属性，这并不能从根本上解决长期依 赖性问题。Transformer 模型基于自注意力机制，并加入了并行注意力机制，这使 得它能够捕获远距离的时序特性，避免了长期依赖。因此，本文结合 LSTM 和 Transformer 的模型特点，构建 trans_LSTM 融合模型，并设计了一个基于深度学 习的股票套利策略，进一步与 LSTM 模型和统计方法协整模型进行了详细的比 较分析。本文第一部分概述了研究的背景意义，概述了国内外的研究进展，并对 本文的研究内容和创新点做了简洁的描述。第二部分对套利的基本知识进行了简 单介绍。第三部分对基于 trans_LSTM 的套利模型设计进行模型和策略的介绍。 第四部分实证分析，通过遍历股票组，基于 trans_LSTM 模型、LSTM 以及协整 理论构建预测模型，对套利组合的价差进行预测。 通过构建完整的 trans_LSTM 融合模型和 LSTM 交易策略的套利模型，就科 创板数据集的结果以及跨市套利结果分别在牛市和熊市两种情况下进行对比分 析，发现了基于传统统计方法模型没有发现的可套利组合，说明基于深度学习方 法的套利问题研究适用的广泛性。显示 trans_LSTM 融合模型的性能更好，得到 更高的收益率，并且符合牛市和熊市回测时间的股票现状，验证了 trans_LSTM 融合模型在股票套利策略中的高效性和适用性，为金融市场中深度学习的进一步 发展提供了有价值的参考。

Lately, the stock securities market in China has seen robust growth. The slow adoption of arbitrage tactics in the progression of quantitative trading, spanning from theoretical studies to institutional structuring and practical investigation, signifies a pivotal moment in the evolution of China's securities market.The rapid expansion of trading information in financial markets has led to a growing array of elements affecting stock values.As a result, there's been an increased need to analyze technical instruments in the financial sector. The progression in computer technology has enabled the widespread adoption of machine learning techniques, with deep learning, a key element of machine learning, offering substantial benefits in managing nonlinear connections and complex, highdimensional data. The Long Short Term Memory Network Model (LSTM) is suitable for dealing with time series related problems due to its excellent recursive properties. However, the LSTM model has its unique structural properties, which cannot fundamentally solve the problem of long-term dependency. The Transformer model is based on self attention mechanism and incorporates parallel attention mechanism, which enables it to capture long-range temporal characteristics and avoid long-term dependencies. Therefore, this article combines the characteristics of LSTM and Transformer models, constructs a trans-LSTM fusion model, and designs a stock arbitrage strategy based on deep learning, further comparing and analyzing it in detail with LSTM models and statistical cointegration models. The first part of this article provides an overview of the background significance of the research, summarizes the research progress at home and abroad, and provides a concise description of the research content and innovation points of this article. The second part provides a brief introduction to the basic knowledge of arbitrage. The third part introduces the model and strategy for designing arbitrage models based on trans_LSTM. The fourth part is empirical analysis, which involves traversing stock groups and constructing a prediction model based on the trans-LSTM model, LSTM, and cointegration theory to predict the price difference of arbitrage portfolios. By constructing a complete trans LSTM fusion model and an arbitrage model of LSTM trading strategy, a comparative analysis was conducted on the results of the Science and Technology Innovation Board dataset and cross market arbitrage results in bull and bear markets, respectively. Arbitrage combinations that were not found in traditional statistical methods were found, indicating the widespread applicability of deep learning based arbitrage research. The performance of the trans-LSTM fusion model is shown to be better, resulting in higher returns and consistent with the stock situation during bull and bear market backtesting times. This validates the efficiency and applicability of the trans-LSTM fusion model in stock arbitrage strategies, providing valuable reference for the further development of deep learning in financial markets.