An examination of some statistical and economic models involving exchange rates.

This dissertation is concerned with the examination of some widely employed nonlinear exchange rate models. In particular, its aim is to assess how well non-linear statistical models accommodate the theoretical implications contained in economic models and how well they are able to capture the empirical properties of the data.Chapter 2 gives a brief background to the concept of PPP and discusses the role of transaction costs in economic models, making it necessary to model exchange rates within a non-linear framework. Parametric as well as non-parametric statistical techniques are applied to a long time-series data set to give an indication of the empirical validity of non-linearity in real exchange rates. Wide threshold bands are found to be a common characteristic of real exchange rate data. Chapter 3 studies the fitness of the ESTAR model for real exchange rate modelling. It is shown that wide threshold bands in the empirical data necessitate a small transition function parameter in the exponential regime weighting function, leading to difficulties in the meaningful interpretation of regimes. When this occurs, it is also shown that the ESTAR model is weakly identified over the range of the sample data that one generally works with. These results are illustrated on an empirical data set by replicating the often cited study of Taylor et al. (2001).In Chapter 4 and Chapter 5 a number of non-linear models are evaluated. Simulation experiments indicate that LM style tests that are commonly employed in the literature to test for ESTAR non-linearity have a very low probability of rejecting the false null hypothesis of linearity when the true data generating process is in fact the ESTAR model of Taylor et al. (2001). It is further shown that, contrary to the claims of the recent study by Rapach and Wohar (2006), long-horizon forecasts from the ESTAR model converge to the unconditional mean of the series, so that there is no gain in utilising the ESTAR model for long-horizon forecasts. Studying the Markov switching model of Bergman and Hansson (2005) reveals that the model does not generate any non-linearity as predicted from economic models