1. 2003 is as far back as I can get data, so it can't go any further than that. A 5-6 year window seems to be a good representation of the calibur of a team. How good a team was 8-10 years prior doesn't have a whole lot of relevance to present day but the period is long enough that one fluky year won't move the needle too much.

2. Strength of schedule is factored into the historical numbers. The numbers used to determine the rankings are all adjusted for strength of opponent in the given year. What is not accounted for is how this season's schedule will affect final records. Boise St and LSU are both essentially +17 but you would be crazy to think that they would have the same record. Boise St's average opponent is a -4.6 rating and LSU's average opponent is +2.0.  The two teams could play equally as well but LSU will likely have a worse record.

The Irish are actually down at #46.  They were +5.2 last year but their historical is a brutal -0.3 thanks to a poor 2003 and brutal 2007-08.

It's not really about the players returning, it's about the program's historical performance. Because of the nature of college football, team success is reasonably consistent over the years as classes come and go. Some obviously do better than others but this is an attempt to establish a baseline of where the program is at. The previous year is used to account for the most current snapshot of program performance.

I am a bit limited on my dataset, but using what I had, here is what I found. As you increase the years used for the historic portion, the total Mean Absolute Deviation decreases. In other words, the more years back you go, the more accurate the prediction is. This obviously has limits, but with the data from 2003-2009, for each year you add back, you reduce your error by 4-5%. At five years you have reduced the error about 20% from just repredicting the previous season. 

So if you have justified using more years (and through 2009 results, using all years possible has been justified) then the question becomes what is the right mix between new and old. For the completed 2009 season, the answer was 65% old, 35% new which I just rounded to 2/3 and 1/3 because the error isn't very sensitive to small changes in splits. This does make a big difference for team's like Michigan. Depending on which way you move the splits can make a big difference for teams whose recent performance significantly varies from historic performance.