What is the most appropriate model for generating scenarios for daily foreign exchange rates?

This paper investigates the most appropriate model for generating scenarios for daily foreign exchange rates for a long history of a large number of daily exchange rates and finds: returns are not normal; a mean reversion model is rarely appropriate; sampling from historical returns (natural log differenced data) will capture the basic features of the mean of the return data but will ignore the autocorrelation in the mean and variance of returns; using a fat-tailed distributional assumption by matching the kurtosis of the historical data will capture the excess kurtosis of the data but similarly ignore these autocorrelations; a GARCH(1,1) model is in most cases sufficient to model time dependence of the conditional variance and will generate returns with excess kurtosis. In some cases an MA(1) - GARCH(1,1) model is required to capture residual autocorrelation, and in a few case more complicated ARMA(p,q) - GARCH(1,1) models are needed.