I typically do link these things, but it's
A Statistical Model that Predicts Listeners' Preference Ratings of In-Ear Headphones, parts
1 and
2 that I'm referencing. Bear in mind, we're talking about a consistent clustering of preferences with response that, at most, can be individually modified by a small portion of the canal's length; essentially the worst-case scenario, from a standpoint that HRTF differences are the major timbral issue with headphones.
That was the methodology used in the paper
@pozz is referencing - am I misunderstanding your contention, here? I was parsing your complaint to be that without knowing the correlation of a headphone's in situ response and the wearer's HRTF, we can't predict subjective response.