Abstract:
|
We provide a refined means of implementing MAPE-R, a modification of "Mean Absolute Percent Error" (MAPE), recently introduced by Swanson, Tayman, and Barr. MAPE-R was found to provide a more meaningful representation of average error than MAPE when extreme errors are present. MAPE-R does this through a two-step process that first uses a non-linear (Box-Cox) transformation that reduces the effect of extreme errors and then re-expresses average error into the original scale. We discuss the key conceptual features of MAPE-R, as well as its mathematical properties, and descibe a procedure used to determine when to use MAPE-R instead of MAPE. We also provide a large-scale empirical comparison of MAPE-R and MAPE by using block group population counts from Census 2000 (n=208,668) to conduct an ex post facto evaluation of the accuracy of forecasts done by CACI, Inc., a commercial vendor of demographic statistics. We conclude: 1.) that the process we suggest for implementing MAPE-R makes it easier to use than the process originally suggested by Swanson, Tayman, and Barr; and 2.) that MAPE-R offers a meaningful representation of average error even when when extreme errors are not present.
|