Week 8 in review

Flattered and undersold

Flattered · the score ran ahead of the play by more than 7 points

• Houston at Utah, 13–21 L-8 vs -23 implied · +15
  The scoreboard flattered Houston: the -8 margin ran 15 points ahead of the -23 implied by per-play quality (PPA (expected points added per play) -0.02 vs +0.36, success rate 41% vs 55%) (Utah turned it over twice).
• Miami vs Pittsburgh, 52–23 W+29 vs +14 implied · +15
  The scoreboard flattered Miami: the +29 margin ran 15 points ahead of the +14 implied by per-play quality (PPA (expected points added per play) +0.37 vs +0.23, success rate 41% vs 40%) (Pittsburgh turned it over once).
• Iowa at Minnesota, 10–13 L-3 vs -17 implied · +14
  The scoreboard flattered Iowa: the -3 margin ran 14 points ahead of the -17 implied by per-play quality (PPA (expected points added per play) -0.31 vs -0.06, though Iowa was more consistent: success rate 34% vs 29%) (Minnesota turned it over once).
• Mississippi State vs Oklahoma, 15–23 L-8 vs -21 implied · +13
  The scoreboard flattered Mississippi State: the -8 margin ran 13 points ahead of the -21 implied by per-play quality (PPA (expected points added per play) -0.15 vs +0.16, success rate 34% vs 33%) (Oklahoma turned it over once).

Undersold · the score ran behind the play by more than 7 points

• Oregon at Illinois, 10–21 L-11 vs +9 implied · -20
  The scoreboard undersold Oregon: the -11 margin ran 20 points behind the +9 implied by per-play quality (PPA (expected points added per play) +0.16 vs +0.06, success rate 46% vs 36%) (Oregon turned it over once).
• Northwestern vs Rutgers, 42–14 W+28 vs +46 implied · -18
  The scoreboard undersold Northwestern: the +28 margin ran 18 points behind the +46 implied by per-play quality (PPA (expected points added per play) +0.46 vs -0.23, success rate 56% vs 35%) (Northwestern turned it over once).
• Alabama vs Texas A&M, 31–20 W+11 vs +20 implied · -9
  The scoreboard undersold Alabama: the +11 margin ran 9 points behind the +20 implied by per-play quality (PPA (expected points added per play) +0.21 vs -0.12, success rate 43% vs 35%) (Alabama turned it over once).
• Boise State at Washington State, 31–10 W+21 vs +30 implied · -9
  The scoreboard undersold Boise State: the +21 margin ran 9 points behind the +30 implied by per-play quality (PPA (expected points added per play) +0.30 vs -0.14, success rate 40% vs 32%).
• Texas Tech at Cincinnati, 24–37 L-13 vs -5 implied · -8
  The scoreboard undersold Texas Tech: the -13 margin ran 8 points behind the -5 implied by per-play quality (PPA (expected points added per play) -0.03 vs +0.04, success rate 40% vs 43%) (Texas Tech turned it over 3 times).

Closest projection

Over the 21 games this week where SP+, Elo and the market all existed before kickoff.

Our model experimental · 21 eligible games · vs SP+: 13.8 against 10.4, closest 4 / 17 · vs Elo: 13.8 against 13.7, closest 10 / 11 · vs Market: 13.8 against 10.9, closest 6 / 15 · all four: our model closest 1 of 21

Our call · 21 games where SP+, Elo and Our model all existed before kickoff, and each model on those same games, the market among them:

Our call against the market, head to head: Our call was at least as close in 8 of 21, the market in 13 (a tie counts for both).

Who's been right: every model

Each model is graded on every game it predicted before kickoff, from the first game of the season. The market is graded like every model.

Market for Weeks 1, 2, 4, 6 (54 games) and Elo for Week 1 (29 games) come from their published pre-game numbers (CollegeFootballData), read after the fact: reconstructed, never a number published after kickoff.

Record (ours) names a winner only: the better record before the week, the home team on equal records. ESPN FPI and SRS come from CollegeFootballData's ratings; Sagarin (his PREDICTOR rating), Sonny Moore and The Margin from their own pages, credited by name on our models page. A game counts only when the model's number was saved before its kickoff.

Season to date through Week 8. The table above is the head to head: the models on the same games this week.

Upsets

Ranked teams that lost to an unranked or lower-ranked opponent, by the AP Top 25 in effect for week 8, with each model's pregame margin from the ranked side, the market among them; ✓ when a model had the winner, ✗ when it did not.

• #2 BYU lost at UCFUCF 39, BYU 31
  SP+ ✗

  +14.1

  Elo ✗

  +9.3

  Market ✗

  +7.5

  Our model ✗

  +12.6
• #7 Ole Miss lost at #19 TexasTexas 34, Ole Miss 27
  SP+ ✓

  -6.5

  Elo ✓

  -0.7

  Market ✓

  -3.0

  Our model ✓

  -3.5
• #8 Texas Tech lost at CincinnatiCincinnati 37, Texas Tech 24
  SP+ ✗

  +14.2

  Elo ✗

  +14.5

  Market ✗

  +11.5

  Our model ✗

  +12.0
• #11 Tennessee lost at South CarolinaSouth Carolina 27, Tennessee 24
  SP+ ✗

  +5.0

  Elo ✗

  +2.8

  Market ✓

  -2.5

  Our model ✗

  +5.2
• #15 Iowa lost at MinnesotaMinnesota 13, Iowa 10
  SP+ ✗

  +8.0

  Elo ✗

  +14.5

  Market ✗

  +5.5

  Our model ✗

  +8.5
• #16 USC lost at WisconsinWisconsin 24, USC 7
  SP+ ✗

  +4.5

  Elo ✗

  +3.4

  Market ✗

  +2.5

  Our model ✗

  +7.5

Called it: SP+ 1 of 6 · Elo 1 of 6 · Market 2 of 6 · Our model 1 of 6.

Biggest swings

The plays that moved win probability most, across the 19 games with win-probability data, from the home side.

• Indiana at MichiganMichigan 67% → 7%
  Q4 0:00, 3rd & 5: (00:00) No Huddle-Shotgun #14 T.Carr pass incomplete short right to #6 J.Buchanan thrown to IU 14 broken up by #22 J.Sharpe · final Michigan 6, Indiana 13
• Duke at VirginiaVirginia 61% → 10%
  Q4 1:20, 2nd & 8: (01:20) Shotgun #20 N.Sheppard pass complete deep right to #18 B.Keefe caught at UVA01, for 28 yards to the UVA00 TOUCHDOWN, clock 01:14, 1ST DOWN #25 G.Strange kick attempt failed (H: #38 A.Zamani, LS: #52 J.Fitzgerald) · final Virginia 16, Duke 19
• Iowa at MinnesotaMinnesota 86% → 39%
  Q4 3:51, 2nd & 9: (03:51) Shotgun #9 H.Brown pass complete deep middle to #2 R.Vander Zee caught at MINN12, for 37 yards to the MINN00 TOUCHDOWN, clock 03:46, 1ST DOWN #92 C.Buhr kick attempt good (H: #35 T.Philpott, LS: #69 I.Speltz) · final Minnesota 13, Iowa 10

Regression watch, checked

Last week's teams whose score ran more than 7 points from the play, and what they did this week against their season-to-date form (average margin before the game, home field removed). The 2022–25 study found flattered teams fell -3.0 short of form the next game and undersold teams beat it by +1.7.

• BYU · flattered +12 vs Notre Dameat UCF 31–39 L · -23.0 vs form (-13.3 vs the field)
• Texas · flattered +11 vs Floridavs Ole Miss 34–27 W · -6.6 vs form (+3.2 vs the field)
• Florida State · flattered +9 vs Miamidid not play
• Utah · flattered +7 vs Coloradovs Houston 21–13 W · -19.0 vs form (-9.3 vs the field)
• Penn State · flattered +7 vs Michigandid not play
• Wake Forest · undersold -14 vs Californiadid not play
• Louisville · undersold -13 vs Syracusedid not play
• Washington · undersold -13 vs Purduedid not play
• Northwestern · undersold -13 vs Michigan Statevs Rutgers 42–14 W · +9.8 vs form (+19.6 vs the field)
• Notre Dame · undersold -12 vs BYUdid not play
• Florida · undersold -11 vs Texasdid not play
• Miami · undersold -9 vs Florida Statevs Pittsburgh 52–23 W · -8.0 vs form (+1.7 vs the field)
• Michigan · undersold -7 vs Penn Statevs Indiana 6–13 L · -14.3 vs form (-4.6 vs the field)

Baseline: all 30 page teams that played ran -9.7 vs form this week. The baseline matters because early-season form comes mostly from easier non-conference games, so the whole field tends to fall short of it once conference play starts; a flagged group is read against the field, not against zero.

Flattered: -16.2 vs form over 3, -6.5 against the baseline (-3.0 in the study) · Undersold: -4.2 vs form over 3, +5.6 against the baseline (+1.7 in the study). One week is a few teams and proves nothing either way; the point is that the check is visible every week.