Can We Beat the Random-Walk Model for the South African Rand–U.S. Dollar and South African Rand–UK Pound Exchange Rates? Evidence from Dynamic Model Averaging

Abstract

Traditionally, the literature on forecasting exchange rates with many potential predictors has primarily only accounted for parameter uncertainty using Bayesian model averaging (BMA). Although BMA-based models of exchange rates tend to outperform the random-walk model, we show that when accounting for model uncertainty over and above parameter uncertainty through the use of dynamic model averaging (DMA) and dynamic model selection (DMS), the gains relative to the random-walk model are even bigger. That is, DMA and DMS models outperform not only the random-walk model, but also the BMA model of exchange rates. Furthermore, sensitivity analysis reveals that in exchange-rate modeling, accounting for parameter uncertainty may even be more important than parameter uncertainty. Our results are based on fifteen potential predictors used to forecast two South African rand–based exchange rates. We also unveil variables, which tend to vary over time, that are good predictors of the rand–dollar and rand–pound exchange rates at different forecasting horizons.

KEY WORDS:

• Bayesian
• exchange rates
• forecasting
• macroeconomic fundamentals
• state-space models

Acknowledgments

The authors thank Ali Kutan (the editor), Dimitris Korobilis, and two anonymous referees for many helpful comments. Any remaining errors are solely the authors’.

Notes

1. One could choose values for λ and α based on forecast performance as in Grassi and Santucci De Magistris (2013), but this would bias our results in favour of DMA and is not a valid procedure for out-of-sample forecasting (Koop and Korobilis 2011). Alternatively, when forecasting at time, we could consider a grid of values for λ and α and select the value that yields the highest value for the marginal likelihood or an information criterion, which essentially amounts to treating λ and α as unknown parameters. However, this would greatly add to the computational burden, so much so that it might be impossible to do forecasting in real time (Koop and Korobilis 2011). Hence, we follow Koop and Korobilis (2011, 2012) and select values for the forgetting factors, but we do carry out a sensitivity analysis in the Empirical Results section.

2. Experimentation with up to four lags of the dependent variable shows that one lag length leads to the best forecast performance for both the rand–pound and rand–dollar exchange rates.