This paper provides a comprehensive framework for comparing predictors of univariate time series in the mean square norm. Initially, the forecast errors are assumed to be unbiased, independent, and normally distributed. Each of these is progressively relaxed. A new heteroscedasticity and autocorrelation consistent statistic for forecast comparison is derived. Finite sample distributions are tabulated in a sequence of Monte Carlo exercises. Power is examined by comparing forecast errors from a moving average model with misspecified autoregressive alternatives.
Mean squared prediction error robust forecast comparison