Bias Correction and Out-of-Sample Forecast Accuracy

We evaluate the usefulness of bias-correction methods for autoregressive (AR) models in terms of out-of-sample forecast accuracy, employing two popular methods proposed by Hansen (1999) and So and Shin (1999). Our Monte Carlo simulations show that these methods do not necessarily achieve better forecasting performances than the bias-uncorrected Least Squares (LS) method, because bias correction tends to increase the variance of the estimator. There is a gain from correcting for bias only when the true data generating process is sufficiently persistent. Though the bias arises in finite samples, the sample size (N) is not a crucial factor of the gains from bias-correction, because both the bias and the variance tend to decrease as N goes up. We also provide a real data application with 7 commodity price indices which confirms our findings.