# Why is MASE scaled by the mean absolute error produced by a naive forecast calculated on the in-sample data

Wouldn't a better scaling factor be with the MAE produced by a naive forecast on the test data itself?

When evaluating MASE for the training set, this essentially becomes a comparison for the forecast model with a naive one, why do we not take this approach with the test set?

• Imagine your test set is a single observation, or two observations that happen to be the same value. – Rob Hyndman Feb 23 '17 at 21:43