Title: Multivariate Bayesian predictive synthesis in macroeconomic forecasting
Authors: Knut Are Aastveit - Norges Bank (Norway) [presenting]
Kenichiro McAlinn - Temple University (United States)
Jouchi Nakajima - Bank for International Settlements (Switzerland)
Mike West - Duke University (United States)
Abstract: The aim is to develop the methodology of Bayesian predictive synthesis (BPS) in multivariate time series forecasting with a detailed application in multi-step macro-economic forecasting. Based on foundations in coherent Bayesian reasoning with predictive and decision analytic goals, BPS defines a methodological framework for evaluation, calibration, comparison, and context- and data-informed combination of multiple forecast densities. The BPS framework naturally allows modeling and estimation-- sequentially and adaptively over time-- of varying forecast biases and facets of miscalibration of individual forecast densities, and critically of time-varying inter-dependencies among models or forecasters over multiple series.BPS analysis is developed in one subset of the implied dynamic multivariate latent factor model. Bayesian simulation-based computation enables implementation. A study of sequential BPS analysis in a multiple macroeconomic time series study with US data highlights the potential improvement of forecasts among all series and forecast horizons, as well as the dynamic relationships among forecasting agents over multiple series.