GIE Report Card: Week 5

The Damage Report

Through five weeks of the 2026 season, GIE has called 56 games and landed on the favorite in 37 of them — a 66.1% accuracy rate that sits in the expected range for a model operating across FBS competition. The real story lives in the variance: the engine's confidence buckets performed almost exactly as they should at the extremes, then collapsed in the middle.

When GIE assigned a 90%-plus win probability, the favorite went 11-0. Perfect. At 75-90%, the record was 12-6, which tracks the model's stated confidence. But the 60-75% bucket — where most games actually live — produced a 10-6 record against an expected 60-75%. That's the margin between a quiet week and a catastrophic one. The 50-60% bucket went 4-7, a result so bad it almost requires an explanation that GIE itself cannot provide.

The misses paint a portrait of a model that saw what it saw and got punished for it. Miami (OH) came in at 90% win probability; Bowling Green won 24-20. Penn State sat at 75%; Northwestern won 34-13. Michigan at 74%; Minnesota 20, Michigan 14. These weren't edge cases. These were games where the pregame positioning was clean, the data supported the call, and the field produced a different answer.

The mid-confidence misses sting differently because they suggest not miscalibration but genuine chaos. When Baylor beat Arizona State as a 57% underdog, that's noise. When Fresno State bludgeoned Washington State 26-6 as a 52% underdog — essentially a coin flip in the model's eyes — something structural either shifted late or escaped the input entirely. The same applies to Georgia Southern over Coastal Carolina at 50-50.

Two patterns emerge from the 19 losses. First: the model's worst blunders came in games where the favorite was projected somewhere between "slight edge" and "clear advantage" — 64% to 83%. GIE nailed the certainties and got humbled by the ambiguities. Second: when the engine was wrong at high confidence (Miami, Penn State, Michigan), the misses were often decisive — not a field goal swing but a two-to-three-possession gap. That suggests not marginal miscalibration but a systematic blindness in specific game contexts.

The 90+ bucket's perfection is real and worth noting, though it also means GIE was conservative early in the season — a model behaving well when it had conviction. The 75-90 bucket's 12-6 record is exactly what you'd expect. It's the 60-75% zone where the model faces its reckoning: in a league of such high variance, even a well-calibrated engine can look foolish when three or four high-leverage plays flip a projection.

Looking Forward

As Week 6 approaches, GIE's early-season record is neither alarming nor confident. What matters next is whether the mid-tier accuracy stabilizes or whether this week's collapse in the 60-75% bucket signals a deeper problem that accumulates over time.