Abstract
Inflation forecasts are highly important in the actual management of monetary policy implying that central banks must acquire accurate inflation forecasts. Given the high risks involved in inflation, models that account for uncertainty are expected to do well in terms of forecasting power. The current paper focuses on estimating 12 different specifications including 10 univariate models and 2 multivariate models. The univariate models are generalized autoregressive conditional heteroskedasticity in mean (GARCH-M) and threshold autoregressive generalized conditional heteroskedasticity in mean (TARCH-M) models assuming three different distributions for the error term as well as an extension of the GARCH-M model that allows for time-varying higher order moments. Furthermore, we employed three models that take into consideration the possibility of structural shifts, namely, Markov switch, threshold autoregressive and time-varying coefficients (TVCs) models. The multivariate models are the vector- half operator model and the dynamic stochastic general equilibrium - vector autoregression model (DSGE-VAR). Results indicate that TVCs model and time-varying higher moments outperform other competing models. Finally, forecasts are improved by generating combined forecasts using equal weights, Bayesian model averaging and dynamic model averaging.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1. For more details about the different extensions of ARCH/GARCH models, see Bollerslev (Citation2009).
2. This section is mainly based on Leon et al. (Citation2005) and their development to the GARCH-type model of skewness and kurtosis.
3. For more details, see Harvey (Citation1989).
4. Details on derivation of the model are explained in Del Negro & Schorfheide (Citation2004).
5. The threshold autoregressive model is estimated in levels given that the inflation series is stationary according to the Phillips–Perron unit root test.
6. Loss function need not be quadratic or even to be symmetric, and forecast errors can be non-Gaussian, non-zero mean, serially correlated and contemporaneously correlated.
7. All Bayesian weights and calculations are estimated by using BMS package inside R software.