基于DBN的金融时序数据建模与决策  

Modeling and Decision-making of Financial Time Series Data with DBN

作　　者：曾志平[1,2] 萧海东[2] 张新鹏[1] 

机构地区：[1]上海大学通信与信息工程学院,上海200072 [2]中国科学院上海高等研究院智慧城市研究中心,上海201210

出　　处：《计算机技术与发展》2017年第4期1-5,共5页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61271349);中国科学院科技服务网络计划(STS计划)项目(KFJ-EW-STS-140);中科院先导项目(XDA06010800);上海市科学技术委员会资助课题(14DZ1119100)

摘　　要：在金融时序数据的分析中经常会遇到一些复杂的非线性系统,利用数学方法很难对这些复杂的系统状态方程准确建模。针对目前金融时序的数据分析复杂性和不确定性等问题,将对复杂非线性系统的模拟转化为对金融时序数据曲线的模式识别,确定了金融时序数据上升、下降以及无规则的各种模式。利用深度学习对非结构化数据处理的优势,提出了一种改进的基于深度信念网络(DBN)决策算法的金融时序数据建模与分析方法。将时序数据转化为非结构化数据,以这些非结构化数据作为深度学习网络的输入层训练DBN金融时序数据模型,应用训练好的模型于金融时序数据样本的预测选取和交易。实验结果表明,利用DBN模型选择的金融数据样本在金融时序数据量化的决策分析中的准确率可达到90.544 2%。In analysis of the financial time series data,some complex nonlinear systems are often encountered. It is difficult to accurately model the state equation of these complex systems with mathematical methods. Faced with the current problem of complexity and uncer- tainty of financial time series analysis, simulations of complex nonlinear systems has been translated into pattern recognition of financial time series data and various patterns of financial time series curves, such as ascending, declining and random, have been determined. By taking use of the advantages of deep learning in unstructured data processing, an improved financial time series modeling and analysis method with improved Deep Belief Network （DBN） decision-making algorithm has been proposed,by which time series data have been transformed into unstructured data to be taken as input of input layer training model for in-depth learning network and to use trained mod- el to predict the financial transaction data sample selection. Experimental results show that the accuracy rate acquired by improved deep belief network method has been achieved by 90.544 2 percent in quantitative analysis of final samples.

关 键 词：深度信念网络 受限玻尔兹曼机 深度学习 金融时序数据 预测与决策 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] F830.59[自动化与计算机技术—计算机科学与技术]