经济周期转折点预测的动态贝叶斯网络分类器模型

摘要

针对现有的经济周期波动转折点预测方法侧重静态函数依赖,或者强调动态序列的时间传递,不能将两方面信息有机结合的情况,给出了经济周期波动转折点预测的动态朴素贝叶斯网络分类器模型,并在此基础上,通过增加隐藏变量层建立了层次动态朴素贝叶斯网络分类器模型,该模型更加灵活、实用和可靠,可广泛用于网络时间序列的预测.%Macroeconomics studies primarily explore the rules of turning points of economic cycles by using the prediction methods of function fitting and time series. However, these two methods are primarily based on the concept of static time series and do not consider dynamic time series. A dynamic Bayesian network combines static and dynamic time series at a given time frame. The network adds the time series function to the existing Bayesian network features, including versatility, efficiency and openness. Dynamic Bayesian networks have been applied to causal analysis, the prediction of multi-variables network time series, and other areas.The dynamic Bayesian network used for prediction is called dynamic Bayesian network classifier. Dynamic naive Bayesian network (DNBN) classifier can be adopted for dynamic prediction. This kind of classifier has the advantages of simplicity and high efficiency.However, on the assumption of conditional independence between attribute variables, the prediction accuracy will be decreased when there is a strong conditional dependency between variables. Gaussian distribution or Gaussian kernel distribution is used for continuous attributes. A large difference exists between actual distribution and Gaussian distribution. Gaussian kernel distribution often has the over-fitting problem that can decrease the classifier's generalization ability.Hidden variables play important roles in time series prediction. The performance of the DNBN classifier can be improved by adding a hidden variable layer. This addition can help establish a dynamic hierarchical Bayesian network (abbrevd. HDNBN )classifier. In a HDNBN classifier, bidden variables have two main functions: ( 1 ) they can use a Mixed Gaussian distribution to replace Gaussian distribution (Mix Ganssian distribution can estimate any distribution of continuous variables). This replacement can increase the reliability of the conditional density estimation and regulate the fitting degree of classifiers by the means and dimensions of hidden variables; and (2) it can aggregate the degree of dependence between attributes. Improved classifiers can be applied to the prediction of network time series in more flexible, effective, reliable and practical manners.