R2 is not a good effect size measure, but none of these arguments should convince anyone. http://data.library.virginia.edu/is-r-squared-useless/ …
-
Another criterion: whether the metric is comparable across models, datasets, range restriction etc. There's likely no one best metric.
And whether it is susceptible to outliers etc. E.g. the squared-family metrics are strongly affected, but e.g. median absolute error is not.