Quantitative Timing Study Based On Dynamic Bayesian Generative Adversarial Networks

As an important part of China's market economy,financial market plays an increasingly important role in economic life.With the rapidly economic growth,the quantitative investment market has ushered in unprecedented prosperity,bringing opportunities and challenges to investors.Financial data has characteristics such as non-linearity and complexity,makes traditional investment strategies difficult to meet people's expectations.Therefore,quantitative investment strategies based on deep learning are gradually emerging.As an important branch of quantitative investment,quantitative timing has increasingly become the mainstream tool for market investment.In the deep learning algorithm,the generative adversarial networks have received great attention.It utilizes the idea of confrontation and has a rigorous mathematical theoretical foundation.It can solve problems that are difficult to other neural network algorithms.The thesis improved a dynamic time-scheduled model of dynamic Bayesian generative adversarial networks based on generative adversarial networks,it solved the problem of mode collapse.Finally,the model verified by the Shanghai-Shenzhen 300 Index from January 2002 to March 2019,then obtained the corresponding quantitative timing strategy.Main tasks as follows:(1)Preprocess the Shanghai-Shenzhen 300 index according to the theoretical basis of quantitative timing,and the preprocessed data is subjected to unit root test to provide preconditions for quantitative timing.The original technical data is used to calculate the relevant technical indicators,and it is used as the candidate feature.The catboost algorithm is combined with the shapley additive explanation value for feature selection,and the feature importance measurement result is obtained.The experimental results show that the model has lower error value and higher credibility than the random forest selected features.(2)Use dynamic Bayesian respectively to optimize the parameters of generator and discriminator of generative adversarial networks,avoid it to fall into local optimum,and solve the problem of easy mode collapse.Then the method uses the model to predict the next-day closing price of the Shanghai-Shenzhen 300 Index according to the characteristics selected by catboost.The model is verified by experiments to be better than the long short term memory and wasserstein generative adversarial networks models.(3)Apply the construction model to the quantitative time-selection field,and make decisions based on the next-day closing price predicted by the model.If the predicted value is greater than(less than)the previous day's closing price,and the next day's price is lower than(above)the previous day's closing price,then buy(sell);if the predicted value does not change,no action is taken.The experimental results show that the model can grasp the trend of the market's rising and falling overall,and thus obtain excess returns.