And how does this skill vary during the season?
For decades we've heard the presentation that polls are biased, especially in preseason, and lazily voted on.
Does empirical data generally bear this out?
In college basketball since 2011, ranked teams are a combined 3933-868 (81.9%) against unranked teams. However, if we break it down month-by-month, we get this:
Note the drop-off from December to January. This is when conference games start, and ranked teams start playing 2-3 quality opponents every week. Overall, the early polls seem to perform just as well as the late polls (considering strength of competition) or better.
In college football, we see the exact same thing. Using just the AP poll, and looking back over the last 3 seasons (this data wasn't as ready and I did it half by hand), we see that polls do slightly better in the first ~4 weeks, during the nonconference schedule, and then do worse once conference play starts. The dropoff is likely due to the increase in level of competition; again, this is ranked teams vs unranked teams:
So, overall, the AP Poll actually does better at the beginning of a season, when most ranked teams play cupcakes.
Source for baseketball, playing around with the month: http://www.sports-reference.com/cbb/play-index/tgl_finder.cgi?request=1&match=career&comp_schl_rk=le&val_schl_rk=25&comp_opp_rk=eq&val_opp_rk=NR&game_type=A&game_result=L&is_range=N&order_by=date_game
For football, I went to this page for each week and copied the table into excel, which automatically extracted wins/losses against unranked teams: http://www.espn.com/college-football/rankings/_/poll/1/week/15/year/2014/seasontype/2
Formula in excel for pulling out the W/L data for anyone who wants to keep going this way: =AND(IF(ISNUMBER(SEARCH(" W or L ",Next Week Column)),1,0),IF(ISNUMBER(SEARCH(" # ",Next Week Column)),1,0))