Bayesian learning in high-dimensional state-space models

Speaker Name: 
Hedibert Lopes
Speaker Title: 
Professor
Speaker Organization: 
Statistics and Econometrics at INSPER
Start Time: 
Thursday, February 21, 2019 - 10:00am
End Time: 
Thursday, February 21, 2019 - 11:15am
Location: 
Simularium, E2-180
Organizer: 
Bruno Sanso

Abstract:

Applied Bayesian Statistics has benefited greatly, over the last three decades or so, from the avalanche of Monte Carlo-based tools for approximate posterior inference in highly complex and structure scientific models. Applications ranging from climatology, environmental and health studies to financial and economical applications, and virtually all areas of science where evidence-based scrutiny are mandatory to validate scientific hypotheses, have in turn profit considerably from such technological explosion. The body of scientific researchers associated with the new Department of Statistics at UCSC is a perfect example of how Statistics in general, and Bayesian Statistics in particular, have been influential across various research areas.

My research history is also an example of the interplay of statistical data analysis, statistical modeling and statistical computing. I will discuss my research on high-dimensional state-space models with particular attention to time-varying covariance learning. In one direction, my co-authors and I deal with the curse of dimensionality via parameter reduction, such as those found in the factor modeling literature. In another direction, we heavily regularize the estimation of parameters, to use a quite popular term these days. In one example, we seek dynamic regularization of state-space components in a large state-space model with thousands of state variables. I will start my talk by reviewing the challenges we faced when dealing with such high-dimensional state-space models, with particular emphasis on the computation of posterior summaries of the various components of these large models. Two or three examples taken from my recent research projects and papers will be used throughout the talk for motivational or illustrative purpose. I finish with a few directions of my future research in these and related areas.

 

References: A few research papers referred to during the talk (can be found at www.hedibert.org)

- Parsimony inducing priors for large scale state-space models

- Efficient Bayesian inference for multivariate factor SV models

- Efficient sampling for Gaussian linear regression with arbitrary priors

- Dynamic sparsity on dynamic regression models

Bio:

Hedibert F. Lopes received a Ph.D. in Statistics and Decision Sciences from the Institute of Statistics and Decision Sciences of Duke University in 2000, and an MSc. in Statistics from the Mathematics Institute of the Federal University of Rio de Janeiro (UFRJ) in 1994. Prior to joining Insper in 2013, he worked for ten years at the University of Chicago as Assistant and Associate Professor of Econometrics and Statistics at the Booth School of Business.

He was Professor of Statistics at the Federal University of Fluminense (UFF) and at UFRJ in 1992-96 and 1996-2003, respectively, and Assistant Researcher at the Institute for Applied Economic Research (IPEA/RJ) in 1991-96. He has lectured various courses in undergraduate, masters and doctorate programs over the last two decades, such Bayesian Econometrics, Computational Statistics and Inference Statistics (doctorate) and Business Statistics (MBA).

He conducts research in Bayesian Statistics, Factorial Analysis, Computational Methods, Time Series and Dynamic Models, Multivariate Stochastic Volatility, Extreme Value Theorem, Particle Filters, Spatial Statistics, Microeconometrics and Macroeconometrics. He has published six books, with another three to be published by 2016 (Wiley, Chapman&Hall and Springer), and over 70 scientific papers and technical reports. He has served as an expert for more than 30 international journals, given 200 lectures and administered 25 mini-courses and tutorials over the last decade.

He is Associate Editor of the Journal of Business and Economic Statistics and of Bayesian Analysis and has published papers in international journals, such as the Journal of the American Statistical Association, Annals of Applied Statistics, Statistical Science, Statistics and Computing, Biometrics, Bayesian Analysis, Journal of Time Series Analysis, Econometric Reviews and Computational Statistics and Data Analysis.