Research On Quantitative Stock Selection Strategy Based On Generative Adversarial Network GAN

In recent years,quantitative strategy research has gradually become a hot topic in the field of stock investment.Although in the past few decades,stock market price forecasting and trading strategy research has always been an important research direction in the financial field,previous studies have not formed a consensus.With the development of artificial intelligence and machine learning technology,more and more studies have applied these technologies to stock market forecasting and trading strategies.Generative adversarial network(GAN)is an unsupervised learning model proposed by Ian Goodfellow in 2014.GAN contains two neural networks : generator network and discriminator network.The generator network generates fake data samples by learning the distribution of real data,while the discriminator network evaluates the authenticity of data samples by classifying real data and fake data.This paper aims to use GAN to predict the rise and fall of stock prices in domestic and foreign markets,and explore its practical application effect in the construction of machine learning quantitative stock selection strategy.In this paper,python 3.7 and supermind quantitative trading platforms are used to obtain financial data and historical stock price data of Shanghai and Shenzhen 300 and Shanghai Composite Index from 2020 / 09 / 01 to 2023 / 03 / 01 from tushare and wind database.GAN network is used to construct multi-factor quantitative stock selection strategy.On the basis of the initial strategy,the initial strategy is optimized by batch gradient descent(BGD)method and neuron parameter adjustment,and the effectiveness of the final optimization strategy is tested.Among them,2020 / 09 / 01-2021 / 09 / 01 is the model learning training sample,and 2021 / 09 / 01-2022 / 09 / 01 is the in-sample backtest interval.The results show that the annualized return rate of the optimized final strategy is 29.25 %,which is much higher than the annualized return rate of the initial strategy 11.31 %.The Sharpe ratio and excess return have also been improved,and the maximum withdrawal level has been reduced.At the same time,its backtesting effect is better than the traditional machine learning model to quantify the stock selection strategy,and the strategy has passed the validity test.In the past multi-factor stock selection and machine learning quantitative strategies,feature labels are usually constructed based on historical data indicators.It is inevitable to encounter over-fitting problems when dealing with large sample data,resulting in poor prediction of stock price fluctuations.The generative adversarial network plays a two-way game through its unique generator and discriminator until the Nash equilibrium is reached,and the accuracy is higher,which can better fit the stock price rise and fall.From the perspective of enterprises,by predicting the stock market for time series data,enterprises can better understand market demand and trends;from the perspective of government decision-making,the government can better understand market fluctuations and influencing factors,formulate corresponding policy measures in time,test the effectiveness of government risk management strategies,help improve the government ’s decision-making ability and ability to cope with crises,and maintain market stability and promote economic development.It is of great significance.