GIE Report Card: Week 9

The 75% Week

GIE called 56 games in Week 9 and landed on 42 favorites, a 75% clip that sits comfortably inside the long-run expectation. More useful than the aggregate, though, is how the engine sorted by confidence. In the 75–90% bucket, it went 17–3. In the 60–75% range, 12–5. The 50–60% toss-ups landed 1–5. The 90%+ tier, where near-certainty lives, went 12–1. Each tier performed within or above its theoretical target, which is the quietest way a model has of saying it did its job.

The miss in the 90%+ bucket still stings: James Madison entered as 96% likely and left Sunbelt Stadium with a 24–21 loss to Troy. When a projection that confident gets overturned, it registers as both a statistical event and a narrative one. GIE saw something in the numbers that the field on Saturday did not. That happens. Once every 25 times it shouldn't, and this was it.

Where the Gaps Lived

The real damage concentrated in the tighter ranges, where the model had to choose sides of 50–50 propositions. The 50–60% bucket went 1–5, a drastic miss. UCF at 50% lost 41–9 to Baylor. North Texas at 46% beat Florida Atlantic 48–41. These are the moments when a model's input distribution matters most—when two teams are genuinely close in the projection, and small variance in any direction flips the outcome. GIE's uncertainty was honest; its execution in those margins was not.

The 60–75% tier showed tighter discipline. Nebraska at 86% lost to Washington 17–10. Ohio State, favored at 75%, fell to USC 30–18. These represent the model seeing real separation—a 11–point gap in win probability—but getting read wrong on the day. Washington and USC both executed in ways the pregame projection underweighted. That's different from randomness in a coin-flip scenario; it's the model being outpaced by execution.

The Pattern

Across the 14 misses, a shape emerges: the engine overestimated chalk. Favorites at the 80%+ level combined to go 24–2, a 92% hit rate that exceeds the theoretical 80–90% band. But that efficiency inverted in the 60–75% range and below. When GIE had to split hairs—when two teams looked genuinely close but not identical—it leaned too hard on the favorite's pedigree or prior rating without accounting for situational leverage the underdog possessed.

South Carolina's 42–25 win over Oklahoma, Kansas State's loss to Colorado, and Clemson's 36–24 routing of Florida State all fit the pattern: GIE awarded the chalk between 60–75% confidence, but each underdog brought schematic or roster-level advantages that weren't sufficiently weighted into the pregame number.

The model will self-correct for these ranges in coming weeks. For now, Week 9 was a reminder that precision in the 50–60% bucket and accuracy in the mid-60s separate a functional model from one that's actually sharp.