Dynamic Bayesian Predictive Synthesis in Time Series Forecasting

TL;DR

This paper introduces a Bayesian framework for combining multiple time series forecasts using dynamic latent factor models, improving accuracy by adapting to changing biases and dependencies among forecasters.

Contribution

It presents a novel dynamic latent factor model for forecast synthesis that adapts to time-varying biases and dependencies, unifying existing pooling methods within a Bayesian approach.

Findings

Enhanced forecast accuracy at multiple horizons

Dynamic relationships among forecast densities identified

Models effectively adapt to changing biases and dependencies

Abstract

We discuss model and forecast combination in time series forecasting. A foundational Bayesian perspective based on agent opinion analysis theory defines a new framework for density forecast combination, and encompasses several existing forecast pooling methods. We develop a novel class of dynamic latent factor models for time series forecast synthesis; simulation-based computation enables implementation. These models can dynamically adapt to time-varying biases, miscalibration and inter-dependencies among multiple models or forecasters. A macroeconomic forecasting study highlights the dynamic relationships among synthesized forecast densities, as well as the potential for improved forecast accuracy at multiple horizons.