دانلود رایگان مقاله بهره قابل پیش بینی تجدیدنظر شده

abstract

Using a large sample of time series, Hill et al. (2015) developed a procedure that aims to predict whether a series is ‘‘forecastable’’; that is, whether the standard deviation of the time series will later prove to be larger than that of the forecast errors. Their analysis is based on forecasting using Holt’s method of exponential smoothing. We show that Holt’s method is the wrong one to use for their time series, and we present a number of other corrections and objections to their analysis. © 2016 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

7. Conclusions

Contrary to Hill et al., Granger and Newbold did not define forecastability as the ratio of the demand variance to the forecast error variance, the Holt method is not double exponential smoothing, the Holt method was not the most accurate method in the M3 Competition, and the best forecasting method by a wide margin for these time series is SES, not the Holt method. We question the need for a forecastability quotient in the first place. The correct way to deal with a time series that is difficult to forecast is to compare the accuracy to a naïve benchmark using the MASE, which is easy to interpret. Values of the MASE greater than one indicate that the forecasts are worse, on average, than fitted onestep-ahead forecasts from the naïve method, and this idea is easy to extend to simulated errors in a test or holdout sample. If no forecasting method that can beat the naïve method can be found, then the naïve method is the best choice.