Abstract
This study is motivated by the dearth of models that provide good out-of-sample fit for exchange rates. That is, current models of exchange rate behaviour are poor predictors of subsequent currency movements. An attempt is made to determine if the relationship between exchange rates and fundamental variables can help explain the more extreme exchange rate movements (distributional switches). Models are developed that relate fundamental economic variables to the resulting estimates based on the mixture of normal probability distributions. Parametric estimation procedures (Logit and Probit) are compared with a semi-parametric technique, maximum score estimation (MSCORE), which is relatively untested in the field of finance. The fundamental variables of these models include information on trade balances, money supply changes, interest rate changes, real economic growth, relative inflation rates and changes in stock market indexes. Classification results favour MSCORE. Implications of results and improvements in methodology are discussed.
Notes
GARCH–EGARCH behaviour of exchange rates is not incorporated into our binary modelling. The use of a GARCH model with a conditional Gaussian marginal distribution is inconsistent with the use of a binary Gausssian classification system. The GARCH model is predicated on the variance changing over time, therefore there will be many distributions. By using a mixture of two normals we are saying there are only two distributions.
Other functions can also be used.