金融学院知识库

  本文将建立在多因子理论模型的基础上利用线性回归模型、LSTM神经网络、LightGBM模型构建量化投资策略。即对市场众多影响股票收益率的因子分别运用传统的IC分析法以及随机森林机器学习算法进行有效因子以及重要因子筛选，将筛选过后的因子作为LSTM神经网络模型和LightGBM模型两种机器学习算法的输入特征，进行模型训练并进行预测股票涨跌，构建投资组合回测得到收益率曲线，并与多因子线性回归选股模型构建的投资组合收益率曲线进行分析对比。

  本文基于2015-01-01至2021-11-30所有A股股票共64个影响股票收益率的因子数据和相应的月收益率进行建模分析，构建六组量化投资策略与一组改进后的量化投资策略，并基于相同的环境回测，研究发现：回测期间内（1）IC分析法和随机森林机器学习算法两种方法都能有效的对初始因子库因子进行筛选，但两种方法筛选后的因子存在较大差异；（2）基于线性回归模型、LightGBM、LSTM模型构建的七组量化投资策略在回测时期内，均获得了超过基准收益率的收益，并都获得较高夏普比率；（3）在基于线性回归、LSTM模型构建的量化投资策略，将IC分析法作为因子筛选方法，获得更好的回测收益。基于LightGBM模型构建的量化投资策略，将随机森林作为因子筛选方法，获得更好的回测收益；（4）将IC分析法作为因子筛选方法，仅基于LSTM模型构建的量化投资策略回测收益与改进后的基于LightGBM构建的量化投资策略回测收益优于基于线性回归模型构建的量化投资策略回测收益；将随机森林方法作为因子筛选方法，基于LSTM和LightGBM机器学习模型构建的量化投资策略回测收益优于线性模型构建的量化投资策略回测收益。较基于线性回归模型构建的量化投资策略，机器学习模型构建的量化投资策略稳定性、风险控制等指标也都有所提升；（5）在七组量化投资策略模型中，将IC分析法作为因子筛选方法，基于LSTM模型构建的量化投资策略模型在回测期间获得了最佳收益，同时，在回测期间也拥有最大的收益回撤。

     This paper will build a quantitative investment strategy using linear regression model, LSTM neural network, and LightGBM model on the basis of multi-factor theory model. That is, the traditional IC analysis method and the stochastic forest machine learning algorithm are used to screen the effective factors and important factors respectively for many factors affecting the stock returns in the market, and the filtered factors are used as the input characteristics of the two machine learning algorithms of the LSTM neural network model and the LightGBM model, the model is trained and the stock rise and fall is predicted, the portfolio back test is constructed to obtain the yield curve, and the portfolio yield curve constructed by the multi-factor linear regression stock selection model is analyzed and compared.

       Based on the modeling and analysis of a total of 64 factor data and corresponding monthly returns of all A-share stocks from 2015-01-01 to 2021-11-30, six sets of quantitative investment strategies and a set of improved quantitative investment strategies are constructed, and based on the same environmental back test, the study finds that the two methods of (1) IC analysis method and random forest machine learning algorithm during the back testing period can effectively filter the initial factor library factors, but there are large differences in the factors after the screening of the two methods ;(2) Seven groups of quantitative investment strategies based on linear regression models, LightGBM, and LSTM models have obtained returns that exceed the benchmark rate of return and all obtained higher Sharpe ratios during the back testing period; (3) In quantitative investment strategies based on linear regression and LSTM models, IC analysis methods are used as factor screening methods to obtain better back testing returns. Based on the lightGBM model, the quantitative investment strategy uses random forest as a factor screening method to obtain better back testing returns; (4) Using ic analysis as a factor screening method, the back testing return of quantitative investment strategy based only on LSTM model and the back testing return of the improved quantitative investment strategy based on LightGBM are better than the back testing return of quantitative investment strategy based on linear regression model; the back testing return of quantitative investment strategy based on LSTM and LightGBM machine learning model is superior to the back testing return of quantitative investment strategy constructed by linear model. Compared with the quantitative investment strategy based on the linear regression model, the quantitative investment strategy stability and risk control indicators constructed by the machine learning model have also been improved; (5) In the seven sets of quantitative investment strategy models, the IC analysis method is used as a factor screening method, and the quantitative investment strategy model based on the LSTM model obtains the best return during the back testing period, and at the same time, it also has the largest draw down during the back testing period.