This paper provides evidence on problems associated with using standard tests for predictive failure when the errors of a linear regression model are not normally distributed. The ability of a simple bootstrap procedure to give a useful degree of control over the significance levels is examined.