Adaptive Variable Selection for Sequential Prediction in Multivariate Dynamic Models

• Activity Number: 340 - SPEED: Bayesian Methods, Part 1
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Isaac Lavine* and Michael Lindon and Mike West
• Companies: Duke University and Tesla and Duke University
• Keywords: Bayesian Forecasting; Decision Theory; Time Series; Variable Selection

Abstract

We tackle the problem of model uncertainty and forecasting in the setting of on-line Bayesian analysis of multivariate time series. We adopt the perspective of a decision-maker with a specific objective function. An adaptive, Bayesian decision-guided approach is developed to select models for this purpose. Forecasts are made within a theoretically motived Bayesian model averaging framework. A simulation study illustrates how decision-guided variable selection differs from likelihood-based Bayesian model averaging. A case study is developed for long-term forecasting of macroeconomic data, where the only predictors are lagged values. The methodology reveals that the incorporation of older lagged values improves long-term macroeconomic forecasting.