The concept of forecastability is a important. See the article https://frepple.com/blog/demand-classification/ to demonstrate how we successfully explain and quantify the forecastability concept.

In my opinion, the problem is not with forecast accuracy benchmarking in principle, but rather with the metrics used to do this benchmarking. Any metric that does not accommodate forecastability is disqualified. Trying to compare MAD, MSE, MAPE, WMAPE, etc are indeed recipes for disaster. Using metrics derived from these, such as FVA are not any better for this purpose. The one deterministic metric that may be useful is MASE, but it is still a metric that measures value-add, not accuracy per se. I have spent over a decade exploring existing metrics and designing new ones that did not have the above limitations. This ultimately led to the creation of the Total Percentile Error (TPE), which CAN be used to perform accuracy benchmarking. Some materials explaining it and working Excel examples can be found here: http://bit.ly/TPError. Also the website link on my comment has an article on LinkedIn on the same error metric.

Naturally, it would need to be more widely adopted before industry benchmarks become available based on it. The FVA formula can (should) also be adjusted to use TPE rather than MAPE as its base metric, since the latter has horribly low correlation to real business value.