While the accuracy of foreign exchange forecasting models continues to receive considerable attention in the academic literature, virtually all recent studies focus on the adequacy of time series models or various econometric approaches to forecasting exchange rates, often with an interest in estimating risk premia. For example, studies by Baille and Bollerslev (1990), Bekaert and Hodrick (1992), Edison (1991), Fang and Kwong (1991-1992), Gerlow and Irwin (1991), Levine (1991), Pope and Peel (1991), and Taylor (1988), all deal exclusively with forecasting models that employ either a times series or econometric approach. None of these studies examines forecasts made by individual forecasters or forecasting services.
While Frankel and Froot (1987) use survey data to test familiar propositions regarding foreign exchange forecasting, their data is limited to median values for panels of forecasters. Similarly, Froot and Ito (1989) rely on median values of forecasts as well. Kwok and Lubecke (1990) compare a single judgmental forecast to two econometric models, but their data do not permit a thorough test of forecast accuracy across forecasters.
Levich (1979) and Levich (1982) both employ data on forecasts of individual forecasters or forecasting services. However, these papers by Levich are now more than a decade old; further, they do not consider cross rate predictions. Focusing only on the value of foreign currencies against the U.S. dollar is subject to a serious limitation. If the forecaster's ideas about the dollar are erroneous, all of the forecasts will be affected and the forecaster will appear to have no forecasting ability. This impression could arise even if the forecaster's ideas about the value of, say, the German mark against the British pound were deadly accurate. Therefore, considering cross rate forecasts provides a fairer assessment of forecast ability, because it frees the evaluation from bias arising from erroneous forecasts of the value of a single currency.
In each issue, Risk magazine features foreign exchange forecasts by major banks in the Euromarkets.' The forecasts are collected on a particular day, and each forecaster offers a forecast (for 3, 6, 9, and 12 month horizons) for the value of the U.S. dollar against the British pound, Japanese yen, German mark, and the Swiss franc. These predictions of dollar values imply six cross rate predictions (pound/yen, pound/mark, pound/franc, yen/mark, yen/franc, and mark/franc). Risk made these quotations available and granted permission to use these data in this study, subject only to the proviso that forecasts not be attributed to the individual banks that made the forecasts. This study employs data from March 1990 to September 1992, covering 26 different forecast dates and giving a total of 89 forecasts for the value of each currency against the dollar. These forecasts imply 534 cross rate forecasts.
Traditionally, currency forecasts have been evaluated by the performance of forecasters against a naive model. This study employs two benchmark models against which the performance of individual forecasters and a consensus forecast may be tested. The first naive forecast is a "no change forecast" in which today's spot exchange rate is taken as a forecast of the spot rate to prevail at the horizon date. As a second naive model, we consider the forward rate quoted on the day of the forecast as a forecast of the spot rate to prevail at the horizon date.(2) Realized spot rates were taken to be the quoted exchange rate found in The Wall Street Journal at the end of the forecast horizon. This study summarizes results of individual and consensus bank forecasts of direct exchange rates and then analyzes cross rate forecasts against these two naive models.
Let the time of the forecast be time zero, so that S sub 0 is the current spot rate, and let S sub t equal the actual spot rate that prevails at the horizon date t. …