A1751
Title: Impulse response estimation in large-scale time series
Authors: Sumanta Basu - Cornell University (United States) [presenting]
Abstract: Impulse response function (IRF) estimation is a canonical problem in multivariate time series. Impulse responses capture how shocks applied to one component of a multivariate dynamical system propagate to its other components over time. In the high-dimensional regime, the majority of existing works focus on a sparse vector autoregressive (VAR) model specification. In practice, however, imposing sparsity constraints directly on the space of impulse responses can provide a more accurate description of the dynamics. A sparse vector moving average (VMA) model specification is adopted to estimate impulse responses in high-dimensional time series. An iterative algorithm is proposed for learning cumulative impulse response functions and is demonstrated for how this can be used to build graphical models for large-scale dynamical systems. Asymptotic analysis of the proposed method is provided, and its advantages are illustrated over competing alternatives using simulated and a real data set from financial economics.