Time-varying Models for Macroeconomic Forecasts

Abstract

This thesis consists of three studies focusing on ways to detect and model time variation among macroeconomic variables. In these three studies, errors with autoregressive moving averages (ARMA), model averaging, and stochastic volatility (SV) are used to investigate the uncertainty and the instability of macroeconomic dynamics. In particular, I expand upon both univariate (autoregressive; AR) and multivariate (vector autoregressive; VAR) time series models.

Chapter 1 provides a general introduction to the research interest of this thesis. Next, Chapter 2 introduces an ARMA component with SV into the unobserved component model. A transformation to a stacked matrix form of the model is conducted for posterior fast simulation. The proposed model is then used to study macroeconomic time series in the United States (US). The proposed new model provides good full-sample simulation for the majority of the macroeconomic variables, and can improve both the point and the interval forecasting performance of these variables across different horizons.

In Chapter 3, I use real-time macroeconomic variables and both time-varying and equal weights with time-varying parameter models to forecast inflation in the US. Three time-varying coefficient models with three specifications of their error terms are studied. The alternative error-term assumptions are errors with a Gaussian distribution, errors with SV, and errors with moving average SV. Both point forecasts and density forecasts suggest that adding variables and allowing time-varying lag length choice can significantly improve forecasting performance. The forecasting performance of the time-varying and equal weights model combination methods show that adding SV can improve density forecasts but not point forecasts.

Finally, in Chapter 4, I employ a time-varying parameter VAR with SV (TVP-VAR-SV) to analyze the dynamics of renewable electricity generation (REG), gross domestic product (GDP) growth, and CO2 emissions. TVP-VAR-SV and other restricted variants are employed for forecasting REG with data from the US. The empirical results suggest that TVP-VAR-SV is suitable for studying the relationship between REG, GDP, and CO2. The forecasting results suggest that VARs with a time-varying volatility specification can perform much better than those without SV, while allowing for time-varying coefficients does not improve forecasting performance. Show more