| In the process of perfecting the asset pricing model in the financial field,scholars often construct new indicators by mining indicators with strong explanatory power to the rate of return or by machine learning.This paper combines the deep learning method with the traditional asset pricing factor model to construct a multi-factor model based on neural network.This paper selects 18 corporate characteristics of 1961 listed companies from January 2002 to December 2021,and divides them into six categories:liquidity,risk,momentum,profit,value and others.The deep learning framework based on Feng et al.(2018)takes the company characteristics as the input of neural network,and predicts the stock index return by constructing depth characteristics,nonlinear weights and depth factors to fit the excess return of stock index.The specific steps are as follows:firstly,a depth factor model based on the data of China’s A-share market is constructed with Fama-French five factors as the benchmark factor.In order to better capture the local and global characteristics of the data,this paper improves the depth factor model,and uses convolution god to replace the feedforward neural network to construct a convolution depth factor model.The depth factor model and convolution depth factor model are used to fit the excess returns of stock indexes(Shanghai and Shenzhen 300 Index,Shanghai Composite Index and Shenzhen Composite Index),and the pricing error and goodness of fit are used to evaluate the advantages and disadvantages of the models.It is found that compared with Fama-French five-factor model,the depth factor model and convolution depth factor model which combine deep learning method with traditional asset pricing model have better explanatory power,and the convolution depth factor model has the best effect.In addition,by comparing the pricing error and goodness of fit of models with different depth factors and hidden layers,it is concluded that the selection of hidden layers and factors should adopt the principle of moderation,and the over-fitting problem is easy to occur if the model structure is too complicated.Secondly,this paper explores which features in the convolution depth factor model play an important role in fitting the excess return of stock index by replacing the model parameters with zero and keeping other features unchanged.The results show that the daily average turnover rate,Amihud illiquidity and trading volume are more important,and among the top ten characteristics of importance,liquidity features account for the largest proportion,so liquidity features are more important for the excess return of stock index in China’s Ashare market.Finally,the convolution depth factor model is used to predict the stock index return,and the prediction effect is compared with Fama-French five factors,support vector regression,extreme gradient lifting tree and long-term memory neural network model.The results show that the pricing error of convolution depth factor model is the smallest,so the combination of convolution neural network and traditional capital asset pricing model has greater advantages in extracting features and better forecasting effect.However,the goodness of fit of convolution depth factor model is less than that of Fama-French five-factor model,which may be due to the nonlinear characteristics of the model extracted by neural network.In addition,through the two-sample T-test,we know that the predictive effect of convolution depth factor model is significantly different from Fama-French five-factor,support vector regression,extreme gradient lifting tree and long-term memory neural network model.Through empirical analysis,this paper proves that the combination of deep learning method and traditional asset pricing factor model has certain advantages,and this method can be applied to the fitting,prediction and quantitative stock selection of other portfolio returns. |
