This is analytical question and is kind of very general. Please focus on the more accurate forecasting method. You obtained various forecast error measures (MAPE, GMRAE, etc) based on the holdout period. These are in turn viewed as forecast accuracy metrics. At the end of the day, your real interest is in how well (accurately) you can forecast the future, i.e. the rest of 2004 in the current example. Question: Does the accuracy in the holdout period necessarily hold in the out-of-sample period (2004:3-2004:12)