Multivariate Dynamic Modeling and Bayesian Decision Analysis for Macroeconomic Policy

Abstract

There is abundant interest in incorporating information from many economic time series to improve forecasts and inform policy. Traditional methods include time-varying vector autoregression models, in which the series are modeled on the lags of all the series in the vector. However, these models become over-parameterized as the number of series increases. Simultaneous graphical dynamic linear models (SGDLMs) are a recent innovation in modeling and computation for large numbers of time series. In SGDLMs, each data series is modeled with its own specialized and limited set of sparse predictors, with simultaneous relationships represented through contemporaneous predictors. At each time point, the posterior distributions of states and volatilities in each of the decoupled univariate models are independently updated for the new observation. Then, the series-specific models are recoupled to account for cross-series dependencies, and decoupled again to continue sequential estimation and forecasting. This provides a flexible and computationally efficient approach for modeling macroeconomic data, and also enables novel order-free structural analyses for policy decision-making.In this dissertation, I apply SGDLMs, along with the hierarchical special case of dynamic dependence network models, to forecast macroeconomic series from the Federal Reserve Economic Data database. I explore methods for selecting the sparse predictors in an automatic data-driven way, which is an important question for practical utility of SGDLMs. These methods include stochastic search and interventions to sequential updating for dynamic variable selection. Furthermore, I link macroeconomic counterfactual forecasting to Bayesian decision analyses for monetary policy decision-making. This includes novel developments in Bayesian decision analysis, including consideration of multi-step ahead forecast paths and a focus on entire forecast distributions of loss in addition to the standard expected loss.