Alpha 套利是一种基于股票现货市场和股指期货两端反向操作，从而构建投资组 合的交易策略，这种交易策略的好处是可以有效规避证券市场的系统性风险，从而在 承担较小风险的情况下获得更加稳定的收益。随着沪深 300 股指期货的推出，我国 股指期货市场得到巨大发展，量化投资技术在我国也发展迅速。在众多量化投资策略 当中，基于机器学习的量化投资策略又是当下最为先进、潮流的一派。 目前基于机器学习的选股策略研究的文章有很多，但是却没有文章将机器学习运 用于 Alpha 套利策略投资组合的构建，本文是首先对于证券的 Alpha 值进行了重新定 义，并在此基础上首次对机器学习运用于 Alpha 套利策略的可行性进行研究。 现有的机器学习选股策略相关文章都是将股票的收益率作为唯一分类特征，训练 学习器，从而将选股策略转化为一个二分类问题。而 Alpha 套利策略效果的好坏取决 于两个方面，一是能否通过对冲有效规避证券市场的共有风险，二是对冲之后投资策 略是否可以取得较高的收益率。这就决定了如果将机器学习运用于 Alpha 套利，决不 能像现在已有的文献一样，仅仅考虑股票的收益率。所以本文的方法是在单因子回归 模型的基础上重新定义了证券的 α 值，并试图运用机器学习模型构建新的证券 α 值定 义下的 Alpha 套利策略投资组合。 在本文的实证部分将基于不同方法选出不同的投资组合，并回测其有效性，根据 不同方法的套利效果的对比对其进行评价，从而从真实市场的角度佐证机器学习在 Alpha 套利策略上的可行性，希望以此对投资者做出投资决策给予一定的帮助。 关键词，Alpha 套利策略，投资组合，机器学习，支持向量机，神经网络 作 者，金 豆 指导教师，岳兴业II The Feasibility Study of Alpha Arbitrage Strategy Based on Machine Learning ABSTRACT Alpha arbitrage is a trading strategy that bases its portfolio on the reverse operation between the stock spot market and the stock future market. The advantage of this strategy lies in effectively avoiding systemic risk of the stock market thus achieving more stable outcome with relatively low overall risk. The stock market of our nation has witnessed great development since the emergence of CSI 300 index future, so is the quantitative investment technique. And machine learning technique is considered most advanced and popular among the quantitative trading strategies. At present, most literature focuses on the stock selection strategies, but research regarding the application of machine learning technique on building Alpha arbitrage portfolio is scarce. This paper first redefines the α value of securities, and then studies the feasibility of applying machine learning to Alpha arbitrage strategy for the first time. The existing articles on machine learning strategy mostly only use rate of return as classification factor, but by training machine learning programs, we can then transfer stock selection strategy into a binary classification problem. The outcome of Alpha arbitrage depends mainly on two aspects, whether we can effectively avoid shared risk of the stock market or obtain a relatively high return through hedging. And that’s why unlike the existing literature, we cannot consider only stock return as classification factor.Therefore, the method of this paper lies in redefining the α value of securities on the basis of single factor regression model, trying to use the machine learning model to construct the portfolio of Alpha arbitrage strategy under the new definition of α value of securities. The empirical part will select different portfolios on the basis of different methods, and back-test their applicability. Then, we will discuss the hedging effects according to the outcome of different methods and justify the applicability of machine learning technique on Alpha arbitrage strategy, in terms of the real market. We sincerely hope that this essay will provide some insights on investment decisions for the investors. KEYWORDS: Alpha Arbitrage Strategy, Investment Portfolio, Machine Learning, Support Vector Machine, Neural Network written by，Jin dou Supervised by，Yue xingye目录 第一章 绪论1 1.1 研究背景和意义 1 1.1.1 研究背景1 1.1.2 研究意义2 1.2 国内外研究现状 2 1.2.1 Alpha 套利研究现状2 1.2.2 神经网络研究现状4 1.2.3 支持向量机研究现状5 1.3 本文创新之处 6 第二章 相关理论9 2.1 ALPHA 套利9 2.2 神经网络 10 2.2.1 神经网络简介10 2.2.2 误差逆传播算法12 2.3 支持向量机 12 2.3.1 线性支持向量机12 2.3.2 非线性支持向量机14 第三章 ALPHA 套利策略和机器学习方法的有机结合 16 3.1 投资组合的构建 16 3.2 通过机器学习确定参数 18 3.3 一般步骤和算法 19 3.4 该方法的不足之处 25 3.4.1 数据的时效性25 3.4.2 模型的泛化性25 第四章 实证分析27 4.1 处理问题的基本流程 27 4.2 数据获取 284.3 数据预处理 29 4.3.1 缺省值的处理29 4.3.2 0-1 标准化处理29 4.4 进行实验 30 4.4.1 基于训练集数据训练学习器30 4.4.2 基于测试集数据检测预测效果31 4.5 结果展示 31 4.5.1 机器学习模型分类效果31 4.5.2 投资组合收益情况32 第五章 总结与展望38 5.1 总结 38 5.2 展望 38