Selecting exchange rate fundamentals by bootstrap

Research shows that the predictive ability of economic fundamentals for exchange rates varies over time; it may be detected in some periods and disappear in others. This paper uses bootstrap-based methods to select time-specific conditioning information for the prediction of exchange rates. By employing measures of the predictive ability over time, along with statistical and economic evaluation criteria, we find that our approach based on pre-selecting and validating fundamentals across bootstrap replications leads to significant forecast improvements and economic gains relative to the random walk. The approach, known as bumping, selects parsimonious models that have out-of-sample predictive power at the one-month horizon; it is found to outperform various alternative methods, including Bayesian, bagging, and standard forecast combinations.