Forecasting and combining competing models of exchange rate determination

Abstract

This article investigates the out-of-sample forecast performance of a set of competing models of exchange rate determination. We compare standard linear models with models that characterize the relationship between exchange rate and the underlying fundamentals by nonlinear dynamics. Linear models tend to outperform at short forecast horizons especially when deviations from long-term equilibrium are small. In contrast, nonlinear models with more elaborate mean-reverting components dominate at longer horizons especially when deviations from long-term equilibrium are large. The results also suggest that combining different forecasting procedures generally produces more accurate forecasts than can be attained from a single model.

Notes

¹ But see the criticism of Faust et al. (2003).

² See Sarno and Taylor (2002) for a comprehensive discussion of competing models of exchange rate determination.

³ See Engel (1994) on the use of Markov switching models for forecasting short-term exchange rate movements. More recently, Kilian and Taylor (2003) found some evidence on exchange rate predictability at horizon of 2 to 3 years by using ESTAR modelling. However, the power of the results decreases when the horizon is shorter.

⁴ See Diebold and Nason (1990); Meese and Rose (1990, 1991); Engel (1994).

⁵ The inflation rate in each country is calculated as the percentage change in the annual CPI inflation rate, i.e. .

⁶ Note that the ESTAR model can be viewed as a generalization of the double-threshold TAR model.

⁷ We also estimated the model allowing for a shift in the mean of the variables. The results we obtained from the two specifications are very similar with respect to the regime classification as well as to the parameter values. As we expected, the differences between the two models mainly consist of the different pattern of the dynamic propagation of a permanent shift in regime. More precisely, in the MSIH model, the expected growth of the variables responds to a transition from one state to another in a smoother way. See Krolzig (1997) on the peculiarity of the two models.

⁸ See, for example, Bates and Granger (1969), Granger and Ramanathan (1984) and Clemen (1989).

⁹ The loss function is defined as d _t  = g(e _1t )² − g(e _2t )² and d _t  = |g(e _1t )| − |g(e _2t )| for MSE and MAE, respectively.

¹⁰ This term is computed by using the Newey–West lag window.

¹¹ See Granger and Newbold (1976) for a detailed description of the test.