So I havent thoroughly followed the latest discussion in this thread, but on the point of listening window vs on-axis being used in the prference formula, I actually made a
rather long post about this when I suggested to MZKM that he calculate an additional score using the listening Window as well the on-axis.
It's a long one, so I'll summarize what was perhaps my main point:
In a way, Olive's preference papers did not truly correlate individual speakers in their totality of use cases with preference. What he did was correlate specific
data acquired from a group of speakers with preference.
It's a subtle but important distinction; the on-axis curve appears to have had higher correlation with preference than the listening window because it was a more accurate reflection of the direct sound given the specific study methodology.
All 70 of the speakers during the listening tests were set up on axis, toed in completely toward the listener, who were all positioned in the central same seat at 10 ft/3m away.
So it's no wonder listening window had a lower correlation in the formula. 30 degrees horizontal and 10 degrees vertical is simply too wide a span for such fixed seating at 10 feet/3m; you will rarely be more than 5 degrees off axis in any direction with such positioning during dedicated listening.
Really, on-axis should've probably been called 'direct axis' or something. The on-axis curve fares best for preference because it is a better representation of the direct sound Harman's test listeners had than the listening window.
But in the real world, we know most people don't toe in their speakers all the way. Devantier 2002 showed about 65 percent of speakers being off axis (20 degrees was most common after on axis). And coaxial speakers and a few other designs might be better represented by averaging.
I wonder if a smaller listening window - say, 5 degrees in every direction - would've yielded even higher correlations in those tests at 10 feet. We'll never know.