A general framework for predicting returns from multiple currency investments

This paper examines the ability of several popular forecasting models to predict future returns on financial assets denominated in the five major reserve currencies. The paper generalizes the AutoRegressive Conditional Heteroskedasticity (ARCH) model by relaxing some of the underlying restrictive assumptions, forming what we call extended-ARCH models. The paper uses this generalization, as well as general versions of the Random Coefficient (RC) model, in order to predict the rates of return. Out-of-sample predictions based on these models are generated and compared with those of a random walk process and a univariate ARCH model. The main implication of the results is that while direct ARCH estimation can improve on forecasts made on the basis of a random walk model, even better forecasts can be made by generalizing the ARCH process by allowing the coefficients to be time-varying. Thus, the paper provides an empirical implementation of recent work that suggests a method for computing optimal multiple currency mean-variance portfolios.