Bayesian Variable Selection for Nowcasting Economic Time Series

• Activity Number: 220
• Type: Invited
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Marketing
• Author(s): Hal R Varian*+
• Companies: Google
• Keywords: Kalman filter ; time series ; nowcasting ; spike and slab prior ; Markov chain Monte Carlo

Abstract

Economic time series are released at regular, but infrequent intervals. "Nowcasting" refers to the use of external signals to maintain an updated set of beliefs about the current value of the series. Because the number of potential signals is very large, there is considerable uncertainty about which signals are relevant predictors. We describe a method of combining Bayesian spike-and-slab priors with structural time series models to manage the nowcasting problem. Different methods are needed when the variable is an instantaneous measurement of a continuously varying quantity (a "stock" variable, like an interest rate) versus a total that accumulates over time (a "flow" variable, like housing starts or auto sales). We illustrate the methods using high-level summaries of Google search data as predictors.