Updated ESPN FPI Projections for Remaining Games

But "90% chance that a candidate will win" is very different from "polling at 90%" (or polling at 10%). Polling at 90% would lead prognosticators to put the odds of winning an election at 99% or 100% (which is sort of what you said at the end, but you're drawing a tenuous link). Also, teams have definitely won after having a 10% chance of winning (see App State v. Texas A&M -900)

I think their math is probably right, because the population is all college games and the historical results of teams considered comparable to these. 

The math is correct, per the FPI model itself.  It provides a "Win Out" probability for regular season + the conference championship.  Assuming Michigan is between a 75% and 80% favorite in the BIG10 title game, that game plus the OSU win probability yields the 3.2% Win Out probability shown on the FPI page.

Statistics are hard. I would know, I’ve never taken it.

There are two assumptions being made here:

(a) the FPI model produces useful results
(b) football games can be modeled as independent trials.

(a) can actually be checked pretty easily:

There are two simple checks you can do.  The first is to count the number of times the favorites win and compare them to the number of times they're expected to win.  Within the scope of a single weekend, (b) should apply almost perfectly.  There's no logical reason to believe that the Michigan / Hawaii result is dependent, in any way, on Utah / Southern Utah.  Therefore, you can calculate the expected number of wins by the favorites by simple addition -- a 90% favorite contributes 0.9 wins, a 51% favorite contributes 0.51 wins, etc.  Add them up and compare to the actual number of favorites who won.  (KenPom does exactly this, BTW, on his page).

The second check assumes (b) applies to some extent over time, but I think it's a fairly reasonable place to start.  Break the percentages down into buckets -- 0-5%, 6-15%, .. .96-100%.  That's 11 buckets (these choices are arbtirary; pick your own). For each game, look at the win probability of each team, and then assign a result into the appropriate bucket.  i.e., Michigan was a 99.7% FPI favorite against Hawaii, so put a W into the 96-100% bucket and an L into the 0-5% bucket.  (Then, realize that these will necessarily be symmetrical and you can ignore all of the buckets less than 50%).

Anyway, add them all up, and you should find that the winning percentage in the 96-100% bucket is about 98%; in the 86-95%, it's about 90%; the 76-85% is about 80%, etc.  That is, when you get to 100 total events in the 76-85% bucket, you should expect there to be approximately 80 wins and 20 losses.

I believe the appropriate significance test afterwards is a χ2 analysis, but it's been a while -- and anyway, we're just looking for broad strokes here.  You should be able to eyeball it and get an idea for whether or not the algorithm works.

As for (b), that's a little trickier, because games are correlated.  For example, if Michigan loses to Rutgers, they are more likely to lose to MSU, because losing to Rutgers presumably means that all of their quarterbacks' arms fell off or something.  Models like FPI deal with this by adjusting probability as new data comes in, but that's not helpful for this analysis.

You could probably reverse-engineer a correlation coefficient, which would allow you to correct the probability calculation.  A high correlation tends to move the distribution toward the extremes.  In the trivial case, where games are completely correlated, then the only possible records for the remainder of the season would be 0-10 or 10-0.

All that said, your example is entirely incorrect.  Your election model makes no sense; if FPI is working properly, about 20% of teams that are 4-1 underdogs via FPI should win.

Any issue here is with the accuracy of the FPI projections, not the probability math based on them.

Penn State had to win on the final play in 2017. Yes, they won, but to say they had no issues isn’t exactly true.

Iowa could be QB'd by a 98 year old WW2 veteran who needs a cane to walk and coached by Brian Ferentz and I'd still be terrified of them in Kinnick.

Especially in a true sophomore's first road start and first start (though obviously not playing time) against a functional secondary.

This feels like wishful thinking to me.  If Michigan goes 11-1 with a loss in Columbus, its best win is going to be.. uh... MSU at home?  PSU at home?

The paper-soft schedule really cuts into Michigan's margin for error.

According to the Farmer's Almanac:

42% chance of snow.

27% chance of more than 2" during The Game.