Research On Stock Price Prediction And Portfolio Based On Deep Neural Networks

Stock price forecasting and portfolio optimization are important branches of stock market research.An effective stock price forecasting model can not only help the government predict the economic situation,but also guide enterprises to make financing plans.A better portfolio optimization model can help investors get higher expected return at the same risk level.In recent years,many scholars have applied classical machine learning models to predict the future trend of stock price,and combined these models with classic portfolio theory to construct portfolio,and achieved good results.Recently,with the rise of deep learning technology,deep neural networks have been successfully applied to natural language processing,speech recognition and image classification.Due to the superior performance of deep neural networks,their applications in financial market has attracted extensive attention of scholars.Therefore,it is worth studying how to use the deep neural networks to predict the stock price effectively.It also urges us to combine the deep neural networks with the classic portfolio theory to construct portfolios.The specific research contents are as follows:(1)The performance of deep neural networks in stock price prediction is studied.On the one hand,the stock price forecasting performance of multiple hidden layers multilayer perceptron and two classical machine learning models(i.e.support vector machine and random forest)are compared,and the effect of sample classification on stock price prediction of multiple hidden layers multilayer perceptron is studied.The experimental results show that multiple hidden layers multilayer perceptron performs the best,and effective sample classification can further improve the predictive performance of multiple hidden layers multilayer perceptron.On the other,the stock price forecasting performance of long short term memory network based on random forest is studied.And,this model is compared with the long short term memory network and long short term memory network based on principal component analysis.The experimental results show that this model is better than the other two models.(2)Prediction based portfolio optimization models based on deep neural networks are researched.Three classical deep neural network models(i.e.multiple hidden layers multilayer perceptron,long short term memory network and convolution neural network)are used to predict individual stock returns,and portfolio optimization models are built by using predictive errors and semi absolute deviation as risk metric.The experimental results show that prediction based portfolio optimization model with multiple hidden layers multilayer perceptron performs the best,and higher desired return is more suitable for this model.After deducting the transaction fee,it is still significantly better than the state-of-art models.(3)Prediction based omega model with return prediction of long short term memory network is proposed.In this model,the long short term memory network is first used to predict the stock returns,then a portfolio optimization model is established by advancing omega model with long short term memory network's predictive results.And,this model is compared with two prediction based omega models and three equal weight portfolios.Experimental results show that this model performs better than other models,and risk return preference has little influence on this model.After deducting the transaction fee,it still significantly outperforms the state-of-art model.