Note: Most of the times and ages in the following speculation are guesswork. Any nitpicks on that basis should be directed to the bit bucket.
He did return to a modified 2015. The thing that makes the plot work for the whole trilogy is the hysteresis effect demonstrated in so understated a fashion by the photo in the first movie. There’s a delay as the timeline adjusts to its altered history, and the effects propagate fairly slowly down the timestream relative to the subjective time of a time traveler.
Marty made his first significant change to the past (interfering with his parents’ meeting) in the early afternoon of his first day in the past; I would estimate the time at around 1:00-2:00 PM. He was unconscious for several hours, but awakened shortly before dinner time. He ran out during supper and went looking for Doc Brown. Sundown on November 5, 1955 at his approximate location would have been at about 5:08 PM, with the end of civil twilight at about 5:35 PM. Since it was full dark, I’d place supper at around 6:30 PM, and Marty’s arrival at the Doc’s place at around 8:00 PM. That means that roughly 6 or 7 hours elapsed between the initial change and Brown’s first look at the photo, which had just begun to change–Marty’s older brother, David, had just begun to disappear. Guessing David’s age to be about 22-23 (he was out of school and had an office job in the modified future), he would have been born around 1962, seven years downstream from the change. That tells us that the changewave covered 7 years in 7 hours; when it reached David’s date of birth, he started fading from everything that happened subsequently. That gives us one data point.
Marty, who was about 16 (and so born in 1969), didn’t begin to vanish until the dance, a week later. That gives us a second data point: the changewave propagated 14 years in 7 days (168 hours).
That gives a slope of 1 year/hour during the initial surge, dropping to .083 year/hour by the end of the week. I would guess (I’m guessing 'cause it’s been too long since I studied calculus) that it’s asymptotically approaching a real-time propagation (i.e. 1 hour/hour). If it’s not asymptotic, eventually the ripples will die out without changing everything, which would play havoc with causality. An interesting side result of this conjecture is that it would theoretically be possible to alter something far enough downstream that the effects wouldn’t catch up before you died of natural causes. You would never experience the modified timeline, but you eventually will have experienced it. But never mind that, never mind that…
The important thing is that even when the propagation is fastest, immediately after the change, it only travels around 1 year/hour. Since it’s going to be at least a year before young Biff can use the almanac, maybe two, the timeline won’t be significantly impacted until a year downstream. Given our approximated slope, that leaves a 1-2 hour window for Old Biff to return to the future and stagger away/vanish, leaving the way clear for our intrepid protagonists to return to the altered 1985.
In short, Old Biff got back to 2015 before the changes reached caught up. That’ll teach you to ask rabid fans a question.
On preview, I see that others have provided simpler versions. When I get the DVDs, maybe I’ll see how far off I am. Bah, I say. I wrote it, I’m posting it. Maybe it’ll teach me to be less long-winded in the future (not likely).