Commentary by J. Scott Armstrong on Fildes et al.: <93>Generalizing about univariate forecasting methods: further empirical evidence,<94>

Fildes, Hibon, Makridakis and Meade (1998), which will be referred to as FHMM, extends two important published papers. The idea of taking findings from each study and testing them against the data used in the other study is a good one. Such replications and extensions are important in the effort to develop useful generalizations and publication of this paper reflects the commitment of International Journal of Forecasting to replication research. In addition the study examines procedures for estimating smoothing parameters, and it evaluates the need for using multiple starting points when evaluating forecasting methods. On the negative side, FHMM does not fully describe the conditions under which one might expect a given extrapolation method to provide more accurate forecasts than competing methods. This limits the generalizability of its findings. In addition, I believe that the FHMM generalizations are even more limited than they might appear at first glance.