Agenda

Jim Griffin (UCL) and Maria Kalli (KCL)

The seminars will be available online as well, via Zoom.
Link Zoom
ID riunione: 812 4764 0154
Passcode: TS61T0

Jim Griffin (University College London) - Some approaches to modelling high-dimensional multivariate time series

Abstract:
There has been an increasing interest in modelling high-dimensional multivariate economic time series. Many models build on the work-horse Vector AutoRegression (VAR) and its time-varying extension to TVP-VAR. These models can provide better forecasts and structural analysis than low-dimensional models (particularly during crisis periods) but the large number of parameters can be challenging both inferentially and computationally. In this talk, I will review two recent approaches. The first is the Tensor VAR (TVAR) model which uses a tensor structure to achieve dimension reduction in the coefficient matrices of the VAR. I will discuss Bayesian inference in these models and an extension to a time-varying parameter model. The second approach considers the time-varying Factor Augmented VAR (FA-VAR) and uses an autoencoder to extract low-dimensional non-linear factors from high-dimensional data. I will discuss how a shrinkage prior using groupings of the variables can lead to identifiable factors and better predictive performance. 

Maria Kalli (King's College London) - Network Modeling of Asynchronous Change-Points in Multivariate Time Series

Abstract:
We introduce a novel Bayesian method for asynchronous change-point detection in multivariate time series. This method allows for change-points to occur earlier in some (leading) series followed, after a short delay, by change-points in some other (lagging) series. Such dynamic dependence structure is common in fields such as seismology and neurology where a latent event such as an earthquake or seizure causes certain sensors to register change-points before others. We model these lead-lag dependencies via a latent directed graph and provide a hierarchical prior for learning the graph’s structure and parameters. Posterior inference is made tractable by modifying particle MCMC methods designed for univariate change-point problems. We apply our method to both simulated and real datasets from the fields of seismology and neurology. In the simulated data, we find that our method outperforms competing methods in settings where the change-point locations are dependent across series. In the real data applications we show that our model can also uncover an interpretable network structure.