Upsets
“Biggest upset” is a phrase people use without a definition, and the definition is the whole argument. Three of them here, scored before anything was ranked, and kept separate — blending them would produce a list nobody could argue with, which sounds good and is the problem.
Final margin minus the points the market spotted the winner. The purest statement that the number was wrong, and it does not care whether anyone was ranked.
What limits this archive
Two of the three measures need a closing line, and line history is dense from 2013 and thin before it. That is a hard boundary, not a soft one — an archive that quietly dropped the seasons it could not price would present the 1980s as having had no upsets. Rows with no line carry only the poll measure, and the columns show which is which.
The lines are sportsbook numbers only. FiveThirtyEight, Nate Silver, numberfire, teamrankings are excluded: they are model sites, and grading a surprise against another model’s prediction measures the disagreement between two models, not a surprise.
Win probability converts the closing spread through a normal margin with σ = 16 points, the same distribution the live win-probability model uses.