Can MAE be reported in percentage? In Wang et al. "Predicting New Workload or CPU Performance by Analyzing Public Datasets" (2019), the MAE is measured in a standardized space (Mean=0, SD=1). This is discussed in 3.4 Metrics, page 9.
Then right after Figure 6 on page 12, all the MAE references in the paper are presented in percentage, e.g. the average MAE is 5.5.%.
Why is MAE reported in percentage? Does it have anything to do with standardization?
 A: I haven't read the paper in detail (though their equation (2), where they claim that the MAE against the mean is the standard deviation, casts some doubt on the statistical expertise involved). If you divide the MAE by the mean of the time series, you can interpret the result as a weighted Mean Absolute Percentage Error (wMAPE; see Kolassa & Schütz, 2007). This may be what the authors calculated.
A: Mean absolute error is defined as an average of the absolute differences between observed values $y_i$ and the predictions for them $\hat y_i$
$$
\frac{\sum_{i=1}^n |y_i - \hat y_i|}{n}
$$
So it is not a percentage of anything.
My guess is that they mean that the error is measured in percentage point units, because their target is the relative runtime (p. 3), i.e. by how many percentage points does the runtime differ from the reference runtime. In such case, you can say that the machine learning model makes predictions that are off by some amount of relative runtime units, hence the percentage points.
But I must say, that I agree with Stephan Kolassa, that it is rather strange that they discuss things like standard deviations or normal distributions while discussing MAE, so I'd take the discussion with a grain of salt.
