On the finite sample behavior of adaptive estimators

We explore the finite sample performance of adaptive estimators in linear time series regression models. Our results show that for samples of only 50 observations, the 90 percent confidence interval of the adaptive estimator is 20 to 50 percent smaller than the corresponding interval for its GLS counterpart across a range of symmetrical distributions. When the assumption of symmetry is relaxed smaller gains are observed. These results are sensitive to the degree of departure from normality and the precision of the measurement exercise. We further observe that the estimated standard errors are biased downward.