統計学輪講 第11回

There is interest in time series modeling in estimating dynamic parameters that incorporate information from many series, but the Bayesian methods can be computationally intensive for higher dimensions. Simultaneous graphical dynamic linear models (SGDLMs) are coupled systems of univariate time series models that can flexibly represent cross-series relationships while employing fast, sequential learning and forecasting algorithms for practical implementation. The estimation of SGDLMs is developed given pre-determined sparse predictor sets for each series, so an area of current research is developing automated and dynamically adaptive ways to select these predictors. At the University of Tokyo, I am collaborating to find novel methods for such variable selection, including applying shrinkage estimates from the dynamic fused lasso model. In this talk, I will discuss the methodology of SGDLMs and present an empirical example from this recent work. ​