Forecast Selection by Conditional Predictive Ability Tests: An Application to the Yen/Dollar Exchange Rate

In this paper, I propose a new method for forecast selection from a pool of many forecasts. My method has two features. The first is the use of the conditional predictive ability test proposed by Giacomini and White [2006]. Second, I construct a measure with two dimensions: "relative usefulness" and "signal predictability". The measure is designed to rank many forecasts in the order of ex-ante forecast accuracy. Therefore, the ranking can be useful not only for selection of a single forecast but also for forecast combinations. I apply the method to the monthly yen/dollar exchange rate. First, I evaluate the performance of base-line forecasting models including a forecast survey of Japanese companies. Second, I show empirically that my method of switching forecasting models reduces forecast errors compared with a single model.