This paper provides exact evaluations of the distribution of the usual coefficient of determination when the regression model's errors follow an AR(1) or MA(1) process. This provides insights into the extent to which this measure of goodness of fit is distorted by such model mis-specification.