Forecasting South African Macroeconomic Data with a Nonlinear DSGE Model

This paper considers the forecasting performance of a nonlinear dynamic stochastic general equilibrium (DSGE) model. The results are compared to a wide selection of competing models, which include a linear DSGE model and a variety of vector autoregressive (VAR) models. The parameters in the VAR models are estimated with classical and Bayesian techniques; where some of the Bayesian models are augmented with stochastic-variable-selection, time-varying parameters, endogenous structural breaks and various forms of prior-shrinkage (which includes the Minnesota prior as well). The structure of the DSGE models follows that of New-Keynesian varieties, which allow for several nominal and real rigidities. The nonlinear DSGE model makes use of the second-order solution method of Schmitt-Grohe and Uribe (2004) and a particle filter to generate values for the unobserved variables. Most of the parameters in the models are estimated using maximum likelihood techniques. The models are applied to South African macroeconomic data, with an initial in-sample period of 1960Q1 to 1999Q4. The models are then estimated recursively, by extending the in-sample period by a quarter, to generate successive forecasts over the out-of-sample period, 2000Q1 to 2011Q4. We find that the forecasting performance of the nonlinear DSGE model is almost always significantly superior to that of it's linear counterpart; particularly over longer forecasting horizons. The nonlinear DSGE model also outperforms the selection of VAR models in most cases.