Extensions In Dynamic Factor Models And Applications In Economics And Finance

In the recent decades,dynamic factor models have been applied to capture cross-sectional dependent features for multivariate data sets from economic and financial applications.In particular,the driven forces underlying these dependent structure are modeled as unobservable factors,which will be utilized in further analysis and fore-casting.However,as the cross-sectional dimension increases,the cross-sectional depen-dent structure gets more complicated.And as the time dimension increases,the true data generating process is more likely to have some time-varying features.In addition,the actual data gathered in the real-time is highly possible to be irregular.Thus,it be-comes necessary to extend the basic dynamic factor model based on empirical interests.Considering the above background,this thesis systematically investigates the dy-namic factor models and their extensions in the literature.Furthermore,combining the empirical problems of 'business cycle measuring in the real-time','the clustering and diversity of local business fluctuations' and 'modeling volatility for high dimensional asset returns' this thesis summaries the features of current 'mixed-frequency dynam-ic factor model','orthogonal multilevel dynamic factor model' and 'factor stochastic volatility model',and discusses the necessary extensions and how to realize them.The main contents and contributions of this thesis include the following aspects.Firstly,from the perspective of macroeconomic data releasing,and better utilizing the timely information within Mixed-Frequency and Ragged-Edge dataset,this thesis develops a Mixed-frequency Markov-switching dynamic factor model for an unbal-anced datasets together with its Bayesian estimation procedure.Monte Carlo simula-tion study shows the Bayesian method outperforms the maximum likelihood method,and using Ragged-Edge can reduce the measure errors of common factor in the real-time.Based on 256 real-time data sets collected on the data releasing dates since 2008,it shows that our model well characterizes China's business cycle since 1992,and it-s estimation is robust and reliable with respect to real-time ragged-edge datasets and GDP data revisions.In addition,there may exist about 2 to 8 month delays in real-time dating business cycle turning points.Furthermore,among all indicators released in turn within every month,the real-time revision impacts of indicators like Industrial Product,Fiscal Taxation on the contemporal business cycle fluctuation are both substantial and reliable,while the impacts of Industry Revenue,Export and Import total amounts are relatively weak.Secondly,to better model the clustering and diversified features of local business cycle fluctuations,this thesis develops a novel extension to current multilevel dynamic factor models.In the one hand,it identifies the regional grouping of provinces simul-taneously with estimating the factors,rather than treating these grouping as given.In the other hand,it allows the multilevel factors to have dynamic interactions.Such ex-tensions make the model more realistic for studying the local clustering and diversity of business-cycle fluctuations across China's provinces.It shows that 31 provinces can be divided into the leading region,the inland region,the overshooting region and the northwestern region.Estimated regional factors show that there are substantial het-erogeneities among regional business-cycle fluctuations especially during "Subprime crisis" and "4-trillion stimulus package" periods.Further evidence shows that due to regional differences,tightening monetary policy might be less efficient in some region,and tremendous local fiscal stimulus programs may not be necessary for some regions even during national recessions.Thus more region-specific packages,and better co-ordinating among agencies would be helpful in smoothing regional fluctuations and stabilizing the whole economy.Thirdly,a new multilevel dynamic factor stochastic volatility model is proposed to study the possibly high dimensional asset returns data set.Firstly,in comparison to the current multilevel dynamic factor models,the proposed model employs a more general 'multilevel' design such that there may exist more than one layer of paral-lel classification standards.This setting fits with the strong correlation among assets from the same industry,district or the same market.Secondly,all these factors and id-iosyncratic errors include a stochastic volatility component,which allows the returns to have time-varying conditional variances and also time-varying conditional correlations.Thirdly,some recent developments of stochastic volatility model estimation algorithms are adopted such that the Bayesian estimation is still affordable with high dimension-al data.In the empirical study,we divide 329 China's A-share stock into 4 different markets,45 sectors and 28 provinces.The common or specific factors are estimated together with their volatility.Variance decomposition shows the influences of markets,industries and districts varies as the market environment changes.In addition,some example is provided to illustrate the application of this model in portfolio optimization and risk managements.