http://projects.fivethirtyeight.com/flights/
http://fivethirtyeight.com/features/...stest-flights/
http://fivethirtyeight.com/features/...r-2015-update/

It is not a surprise that delays are worst at airports like LGA, JFK, EWR, and ORD, and less of a problem at airports like HNL, PDX, and SAN.

Unlike the vast majority of what 538 does, I found this one not very useful/actionable. VIrgin's average flight is 4 minutes faster than US. When I have a choice between them, between the same city pairs, leaving at about the same time to meet my schedule, with the same number of stops, I might consider this. Which will be about never. (and actually, I wouldn't the variance around the 4 minute mean makes the influence on individual flights pretty close to zero)

I do this kind of analysis for work, and yeah, this 538 bit makes for some good reading, but is useless in practice.

As you note, how many non-stop routes are there where there is any competition? Yeah there's a few, but only a few.

As you further note, four minutes is so little that its useless. You'll lose that sitting in the back of someone else's bus on which you don't have elite status.

The really interesting analytic bits come on head to head routes when you're departing on the odd-man-out airline at someone else's hub. Case in point: I once flew from IAD-CVG on DL (er Comair I think) and we underblocked by 45 minutes. But this is a flight departing a United hub at 3pm, which avoids the massive crush at 5pm. This same flight departing at 5pm could very easily get stuck in a queue 20 deep and chew up massive amounts delay on the ground.

Give all of this, it can be kind of obvious that this is all schedule dependent, down to the exact minute flight(s) are scheduled to leave the gate, both yours and your competitors.

TL;DR: I totally agree with you for two reasons: 1) The opportunities where there's even much of a choice to capitalize on this are few and far between, and 2) So much of this is schedule dependent, and schedules always change, so I'm not sure where the predictive value is.