Risk related non linearities in exchange rates: A comparison of parametric and semiparametric estimates

This paper uses semiparametric techniques to estimate a model of exchange rate determination and compare it to a parametric LSTAR specification. In both cases the nonlinearities are modeled as part of the conditional mean of the process, rather than of its variance. Using a panel dataset for five East European countries for years 1993 - 2001, it results that the non parametric data-driven estimates perform a little better but actually support the LSTAR specification. The dependence of current on lagged exchange rates is confirmed to be non linear, with marginal effects that become very significant and negative for abnormal values of the lagged variable. The PPP hypothesis, is not rejected by the data.