Title: Bayesian dynamic regularised forecast combinations
Authors: Marcel Scharth - The University of Sydney Business School (Australia) [presenting]
Andrey Vasnev - University of Sydney (Australia)
Haonan Zhang - The University of Sydney (Australia)
Abstract: How can we leverage the diversity of professional forecasts and plausible models available for many forecasting problems? Multiple empirical studies have documented that while pooling forecasts improves accuracy in many applications, it is often challenging to design combination schemes that outperform a simple average of the available forecasts--an empirical pattern known as the forecasting combination puzzle. We address this problem by developing a Bayesian approach that automatically computes regularised forecast combinations based on promising subsets of forecasters or models. We formulate this approach as a state space model that additionally accounts for the potential dynamic features of the data, such as autocorrelated errors, heteroscedasticity, attrition of forecasters, and time-varying performance. We propose an estimation algorithm based on Sequential Monte Carlo (SMC). Empirical results illustrate that our approach can improve accuracy relative to simple average combinations and other recently proposed methods based on subset selection and regularisation.