Font Size: a A A

Research On Optimal Design And Stock Price Trend Prediction Based On Deep Belief Network

Posted on:2022-11-23Degree:MasterType:Thesis
Country:ChinaCandidate:S Y XiFull Text:PDF
GTID:2518306746968819Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
As a country's economic barometer,the stock market affects the formulation of national macroeconomic policies and resource allocation to a certain extent.At the same time,it also affects individual investment decisions.In the stock market,investors are most concerned about how to minimize risks and maximize benefits when buying and selling stocks.Therefore,it is very important for investors to grasp the ups and downs of the stock market.The main work of this paper is to study the prediction algorithm strategy of this rising and falling trend.At the same time,based on the general considerations of this research algorithm strategy,the experiment explores the prediction performance of this research strategy on energy load data similar to stock price fluctuation data.As we all know,stock price volatility is easily affected by many factors such as policy and market economy,and the data is highly nonlinear and noisy,which makes it difficult for us to accurately predict it.Over the years,there have been many algorithm strategies for stock data prediction.Some literature studies have shown that Deep Belief Network(DBN)can efficiently extract the features of multi-dimensional data and fit nonlinear data well.Therefore,this paper chooses the DBN network as the research object.Aiming at the problems of the original DBN model in stock forecasts,such as poor ability to process real-valued data,undetermined network structure,and random initialization of network parameters,this paper studies and analyzes the aspects of adding Gaussian noise,automatic adjustment of network structure,and optimization of network parameters.Improve.In order to verify the effectiveness of the improvement in this study,this paper selects two groups of historical stock transaction records for verification and analysis.The results show that the error MAPE values of the two groups of data using the improved model for prediction have decreased by 21% and19.7% respectively,indicating that the improved method in this paper is effective.of.At the same time,this paper also uses this model to verify and predict the gas load data,and the results also prove the effectiveness of the improvement.Therefore,the research results have certain generality and practical value.The main innovations and work of this paper are as follows:(1)The characteristics of stock data are analyzed,the difficulty and feasibility of stock market forecasting are expounded,and the advantages of deep network in stock forecasting are proved.The relevant domestic and foreign research literatures are reviewed to provide theoretical basis for the innovation of this paper.(1)The characteristics of stock data are analyzed,the difficulty and feasibility of stock market forecasting are expounded,and the advantages of deep network in stock forecasting are proved.The relevant domestic and foreign research literatures are reviewed to provide theoretical basis for the innovation of this paper.(2)Using the deep network DBN that can directly process the original data and deeply mine the internal information of the data,analyze and predict the stock price trend.Firstly,the basic idea and model structure of DBN are summarized,and the training process of bottom-up layer-by-layer training and top-down reverse fine-tuning of DBN network is studied;To solve the problem of processing continuous data,Gaussian noise is added to the restricted Boltzmann machine;then different DBN network structures are set up through artificial experience,and the influence of different network structures on the prediction results is discussed and analyzed.(3)Aiming at the disadvantages of artificial experience setting of DBN network parameters,a dynamic Gaussian DBN prediction model(Dynamic Gauss Deep Belief Network,DGDBN)is proposed to improve the efficiency of the network model construction and the accuracy of DBN network prediction.The connection weight quadratic paradigm and the average percentage error are introduced into the deep belief network.First,by calculating and analyzing the square root of the connection weight quadratic paradigm during the network training process,the neurons with low or high weight output are deleted.or split;secondly,the number of hidden layers of the network is determined according to the average percentage error of the network prediction results.It not only solves the problem of determining the network structure but also solves the drawbacks caused by human factors to a certain extent.(4)Based on the dynamic Gaussian DBN model,the Chimp Optimization Algorithm(Ch OA)is further used to improve the DGDBN network,and a dynamic Gaussian deep belief network model based on the Chimp Optimization Algorithm(Chimp Optimization Algorithm-Dynamic Deep Belief Network,Ch OA-DGDBN)is proposed.Firstly,the specific optimization concept of the chimpanzee optimization algorithm is summarized and discussed,and the three standard functions of Sphere,Rastrgin,and Ackley are simulated and modeled by Ch OA,particle swarm optimization and gray wolf optimization algorithm,and the global optimization ability of the Ch OA algorithm is compared and verified;Then,the algorithm is used to optimize the weights,biases and the number of neurons in the initial hidden layer of the DGDBN network;finally,the Ch OA-DGDBN prediction model optimized by the chimpanzee algorithm is used to predict the stock price trend,and the effectiveness of the improved model is verified by comparative experiments.(5)In order to further verify the practicability of the final proposed Ch OADGDBN model,this paper uses it in the actual application prediction scenario of gas load,and discusses and analyzes the experimental results.
Keywords/Search Tags:deep belief network, Gaussian noise, stock trend prediction, dynamic network structure, chimpanzee optimization algorithm
PDF Full Text Request
Related items