A critique of modelling and estimating the effects of consistent conditional variance expectations on exchange rates
A number of authors have recently shown that exchange rate expectations are soundly based on the random walk hypothesis and it is well known the model predicts as good, if not better, than most economic models of exchange rates. This paper suspects that the conditional heteroskedasticity detected in prediction errors of many of these models is primarily due to the reliance on linear first moment, conditional mean representations. In contrast, empirical procedures like GARCH have increased in sophistication to test and correct for the widespread observance of serially correlated second moment, conditional variances in the predictions.
A simple expected utility model which specifies conditional time varying variances is developed to illustrative how empirically correcting for them is sub optimal research strategy. Simply put, if there remains systematic errors in the model's predictions (other than from trading day, seasonality or similar causes) then it is poor research methodology to empirically remove them. The imbalance in terms of the relative lack of modelling higher order moments and the relatively heavy reliance on removing these effects empirically, needs to be addressed.