Skip to content
ML Academy

Introduced on day 16

MASE (mean absolute scaled error)

Mean absolute forecast error divided by the in-sample mean absolute error of a naive (or seasonal naive) forecast; below 1 beats that naive method.

In detail

MASE = mean|e| / ((1/(T-m)) x sum over t=m+1..T of |y_t - y_(t-m)|), with m = 1 for non-seasonal data. The scale comes only from the training series, so it is unit-free and defined when actuals are zero. It is undefined when the training series is constant at lag m. Proposed in Hyndman and Koehler (2006).