Empirical Asset Pricing Research Based On Machine Learnin

Asset pricing plays an important role in improving market cognition.From the perspective of investors,an appropriate prediction model is conducive to the construction of effective portfolios;For managers,market value management and refinancing need the support of effective pricing model;For regulators,effective financial risk measurement is the core of building financial security.This paper combines machine learning with asset pricing research,solving the problem of too little information in traditional multi-factor model.We also break through the constraints of single information set,construct a high-dimensional conditional deep learning factor model,enhance the effectiveness of the model from the perspective of big data and model paradigm,and improve the pricing and forecasting ability.At the same time,considering the asymmetric problem of return and risk in China’s stock market,we also use machine learning to measure the systemic risk,which alleviates the market anomaly.Finally,from the interpretability of machine learning,we discuss the economic mechanism of asset pricing model based on machine learning,including micro and macro perspectives.At first,we analyze risk-return asymmetry in China’s stock market from the perspective of beta anomaly.By constructing the conditional CAPM model based on Artificial Intelligence combined with 666 macro and micro date,we measure the systematic risk more flexibly and intelligently.The empirical results show that the unconditional CAPM model based on monthly data cannot fully explain the risk compensation.After using the conditional CAPM model based on machine learning,the pricing error is significantly reduced.It is also found that the unconditional FF3 model cannot fully explain the beta anomaly and the conditional FF3 model based on machine learning can alleviate the pricing error.In the fifth chapter,we study the return forecasting in China’s stock market based on machine learning in the fourth chapter.Combined with China’s A-share market return and characteristic data,we use various machine learning models,including Lasso,Ridge,Elastic Network,Principal Component Analysis,Partial Least Squares,Random Forest,Gradient Boosting Decision Tree,Neural Network and so on to conduct multi-factor model.The results show that compared with the least square regression model,machine learning overcomes the problems of multicollinearity and over fitting in high-dimensional data.The nonlinear models have better prediction ability than the linear model,mainly because the nonlinear information between variables also involves the pricing power.Among of all models,the Generative Adversarial Network model-GAN performs the best.GAN’s unique dynamic learning function makes it relatively more stable in the case of market fluctuations.The monthly return of the long short portfolio constructed based on the prediction results of GAN reaches 1.13%,and the sharp ratio is 0.71.At last,we study the interpretability of machine learning in finance.In the sixth chapter,we first consider the change of portfolio return in the presence of transaction friction,and find that the portfolio returns based on the nonlinear model are still significant after introducing 0.125% and 0.25% unilateral transaction fee.In the aspect of factor importance,we rank the characteristics based on the empirical results of Gan model,and find that there are three important characteristics that affect the stock return in China’s market:(1)the trend of price and trading volume;(2)Liquidity indicators;(3)Fundamental indicators.The top 10 important factors contribute about 40% pricing abilities of all factors.After that,we use three kinds of indicators such as trading friction,volatility and uncertainty,turnover and liquidity to study whether GAN has the same forecasting ability for different micro characteristic stocks.It is found that Gan model has higher prediction accuracy for stocks with low volatility and high liquidity.In order to further understand the relationship between the asset pricing model and macroeconomic,we also construct a regression model with macroeconomic indicators as dummy variables from the aspects of macroeconomic activity,uncertainty of economy and market,and market sentiment,and find that the pricing power of Gan model is significantly different in business cycles.When the macroeconomy is in the high level of new fixed asset investment,total retail sales of social consumer goods,social financing scale and rediscount interest rate,it indicates that the macro-economy may be in an overheated or irrational state of investors,and the return of deep learning factor will decline.When the market volatility,volume of foreign trade goods,CPI and consumer satisfaction are high,China’s economy and market state are more active and the uncertainty of the external world is higher,and the deep learning factor can obtain more significant excess returns.The main contributions of this paper are as follows:(1)the application of multi class machine learning model in China’s stock market enriches the empirical asset pricing research;(2)We construct a new deep learning model considering the dynamic characteristics of China’s stock market,which improves the market efficiency;(3)We also use machine learning to measure market systemic risk,that strengthens the understanding of financial market risk;(4)The economic ground behind machine learning is explored from the micro and macro perspectives,which improves the interpretability of deep learning model in financial application.