What I was looking for
Whether forecast ambiguity tracks error the same way across data sources.
What I found
The pooled number gets quoted; the split number gets acted on. The first can hide segments running the other way, and the second can be a false positive from trying many splits.
In my forecasting work, a pooled correlation of −0.59 split into +0.175, −0.005, −0.266 and −0.500 by source. Because the splits were registered first, the difference between sources counts as a finding.
The move
I register the splits before the data arrives, so the segments are a result, not a search.