Modeling Regimes with Extremes: The Bayesdfa Package for Identifying and Forecasting Common Trends and Anomalies in Multivariate Time-Series Data

Authors

Eric J. Ward, National Oceanic and Atmospheric Administration
Sean C. Anderson, Fisheries and Oceans Canada
Luis A. Damiano, Iowa State University
Mary E. Hunsicker, National Oceanic and Atmospheric Administration
Michael A. Litzow, University Alaska Fairbanks

Date of this Version

12-2019

Document Type

Article

Citation

The R Journal (December 2019) 11(2); Editor: Michael J. Kane

Comments

Copyright 2019, The R Foundation. Open access material. License: CC BY 4.0 International

Abstract

The bayesdfa package provides a flexible Bayesian modeling framework for applying dynamic factor analysis (DFA) to multivariate time-series data as a dimension reduction tool. The core estimation is done with the Stan probabilistic programming language. In addition to being one of the few Bayesian implementations of DFA, novel features of this model include (1) optionally modeling latent process deviations as drawn from a Student-t distribution to better model extremes, and (2) optionally including autoregressive and moving-average components in the latent trends. Besides estimation, we provide a series of plotting functions to visualize trends, loadings, and model predicted values. A secondary analysis for some applications is to identify regimes in latent trends. We provide a flexible Bayesian implementation of a Hidden Markov Model — also written with Stan — to characterize regime shifts in latent processes. We provide simulation testing and details on parameter sensitivities in supplementary information.