Research On Factor Selection Strategy Based On LSTM Neural Network Model

The theory of quantitative investment has been proposed for nearly seventy years,and the strategy of quantitative investment has undergone a series of evolutions.Nowadays,with the popularity of high-performance computers,artificial intelligence has begun to exert its power in various fields.In recent years,the machine learning field has also made major technical breakthroughs in deep neural networks and applied to the field of quantitative investment.In the field of quantitative investment,the exploration of the application of deep learning models is also gradual.From the Artificial Neural Network(ANN)to the Recurrent Neural Network model(RNN)to the Long Short-Term Memory model(LSTM),the theoretical model solves various problems such as introducing time series,long-term memory training,and mitigating gradient explosion.In this paper,based on the LSTM neural network model,the selected factors are trained and synthesized,and the synthesized factors are used for quantitative investment.In terms of research methods,this paper draws on the theory of factor selection model,and selects 70 different factors such as valuation factors,growth factors and financial quality factors to train and synthesize factors.In terms of factor processing,this paper uses winsorization,neutralization,and standardization to eliminate abnormal data,ensuring the smoothness of the data.In the analysis of the model results,the practicability of the correct rate change,AUC change and loss rate change verification model is adopted,which proves that the model has a certain degree of discrimination.Innovatively,due to the inherent "black box" attribute of the neural network model,it has been criticized for its interpretability.In the environment of the Uqer platform,this paper uses the Python Keras package to try to explain some of the factors,trying to open the "black box" to find out what it is,and has been well explained.In the construction and back testing of the quantitative stock selection model,this paper uses the factor stratification test method to study and demonstrate the differentiation degree of synthetic factors.Finally,we use the repeatedly confirmed factors to quantify the construction of the investment stock selection strategy.In the process of backtesting,the method of comparative argumentation is applied to compare the nonlinear neural network model with the traditional linear statistical model,which shows that the neural network model has a good return on investment.Throughout the process,we can draw the following conclusions: First,the factor which is trained by Long Short-Term Memory model(LSTM)has a good degree of discrimination.Second,the deep neural network model can be explained by the Python Keras package.Third,it is feasible to use the neural network model to train,synthesize,predict,and invest in factors.Fourth,Long Short-Term Memory model has better predictive power and investment ability than traditional statistical models.By using the A-share data from 2007 to 2018,this paper establishes the LSTM factor training and synthesis model,the LSTM factor evaluation and the 5-digit analysis model,and the LSTM factor quantitative investment model.A tentative analysis of model interpretability is done through the Python Keras package.A comprehensive demonstration of the long short-term memory model LSTM in the neural network model has a good performance in quantitative investment.However,there are still some problems in the process of argumentation for further study.For example,the Recurrent Neural Network requires more parameters.The training speed and adjustment of the model are slower and depend on the performance of the computer.The small amount of data is not conducive to the training and play of the model.