Research And Realization Of Time Series Stock Pricing Algorithm

Stock price forecasting has always been one of the core topics in the financial field.Traditional asset pricing methods construct linear relationships to predict future stock prices,but these relationships are difficult to fit into the complex,volatile and frequently fluctuating stock market.Thus,a nonlinear pricing model with Nested Long Short-Term Memory network and autoregressive modules is proposed in this thesis.By processing large-scale time series data through deep learning methods,the nonlinear relationship between stock data features and returns is explored to achieve accurate stock pricing in order to obtain excess returns.In this thesis the shortcomings of current stock pricing methods are first analyzed to understand the problems of inaccurate market fitting by traditional linear technique,inadequate performance of various machine learning methods,and inconsistent research processes.In response to the above problems,the following three improvements are proposed: Firstly,a nonlinear pricing model based on Nested Long Short-Term Memory networks is proposed.The application of this network to capture the latent factors in stock data is able to tap information linkages at a deeper and broader time series dimension compared to traditional recurrent neural networks.Secondly,an autoregressive module is introduced on the basis of the above model.Improvements are made to the problem that Recurrent Neural Network is not sensitive enough to the input scale,so that the model can be better adapted to the stock data,so as to obtain better prediction performance.Thirdly,the Huber loss function is applied to provide more robust performance.In this thesis,Mean Square Error commonly used in regression problems is discarded,and the objective function can be fitted more stably for stock data with high noise and multiple outliers.Furthermore,a complete and standardized experimental procedure is designed in this thesis for the research on stock pricing.For the deep learning model,experiments are designed from the perspectives of predictive and economic performance.Model predictive performance is evaluated using multiple indices and the Diebold-Mariano test.A portfolio based on the model's prediction results is constructed to assess its economic performance.To address the problem that factors are frequently invented by academia but difficult to be tested with all traditional methods,the peculiarities of deep learning are used to evaluate the importance of factors by calculating the influence of the features on the final prediction performance in order to achieve factor screening.Finally,based on the above nonlinear pricing model,an asset allocation intelligent recommendation system based on a micro-service architecture is designed and implemented in this thesis,which implements basic user management functions,and is able to perform intelligent investment portfolio recommendation through the prediction results given by the model.In addition,the system also provides the function of backtesting custom portfolios.In summary,a nonlinear model based on Nested Long Short-Term Memory network is proposed in this thesis,and the prediction performance surpasses traditional and machine learning methods on experimental dataset,and it also shows excellent economic performance after backtesting,indicating that the nonlinear model can achieve more accurate stock pricing.Secondly,the asset allocation intelligent recommendation system based on the microservice structure designed in this thesis also has good practicability.By combining the deep learning method with the financial field,it can be used as a futureoriented investment and financing tool to promote the further development of financial technology.