Modélisation et prévision du taux de change réel effectif américain

This study describes a simple model for predicting the real U.S. exchange rate. Starting with a large number of error-correction models, the authors choose the one giving the best out-of-sample forecasts over the period 1992Q3 - 2002Q1. In the selected model, the effective real exchange rate is cointegrated with relative productivity and the real price of oil. The short-term dynamics depend upon the evolution of the difference in GDP growth rates, the first difference of the ratio of net foreign assets to GDP, the real interest rate differential, and shocks that have a temporary effect on the real price of oil and relative productivity. Out-of-sample forecasts reveal that the model generates mean-squared errors that are systematically and statistically much lower than those from a random-walk or an autoregressive model. This result is largely due to the great stability of the parameters of the cointegration relationship.