Empirical Study On Multi-Factor Stock Picking Model Based On LSTM Neural Network

In recent years,AI has been developing rapidly.The AI robot AlphaGO has won a great victory against human beings.‘Intelligence +' is first written in the government work report of the State Council in 2019.All kinds of evidences show that artificial intelligence plays an increasingly important role in our lives.Quantitative investment is a huge branch of the financial field,in which multi-factor stock selection is a relatively mature stock picking technology.Therefore,in this paper,deep learning is applied to multi-factor stock selection,and a multi-factor stock selection model based on LSTM neural network algorithm is constructed.By comparing with the SVM algorithm,it is found that the LSTM neural network algorithm is more suitable for time series data such as stock than the SVM algorithm.The stock forecasted by LSTM is tested and the result shows that it can obtain more than the benchmark return rate.This paper chooses the factor cross-sectional data of the dynamic Shanghai and Shenzhen 300 component stocks on the last trading day of each week from January2012 to December 2018 as the data sample.Among them,the data from January 2012 to December 2017 are used as training data and validation data of the model.Sample data from January 2018 to December 2018 were used as test data for the model.In the selection of candidate factors,we selected 10 categories of factors,including quality,momentum,value,commonly used technical indicators,per-share indicators,emotions,growth,analyst expectations,basic subjects and derivatives,and earnings and risks,with a total factor of 244.This enlarges the selection range of candidate factors in both breadth and depth.The construction of this model can be divided into the following steps: Firstly,data preprocessing,labeling and principal component analysis are used to reduce the dimension of the original data.Then,the new data are fed into the support vector machine model and the long-term and short-term memory neural network model for comparative analysis.By comparison,it is found that LSTM model is more suitable for non-linear time series data such as stocks.Secondly,training and learning are carried out in LSTM model,and a new model is obtained through learning.Finally,the test data are fed into the new learning model for prediction.Based on the results of the model prediction,some equi-weighted portfolios are selected to test the return,in order to obtain a yield higher than the benchmark of the same period.By evaluating the model with the indicators of annualreturn rate,Sharp ratio,information ratio and maximum retest,it is found that the annual return rate built in this paper exceeds the benchmark return rate in the year of2018.