This was a dark day for the model. A 50% straight-up accuracy — half right, half wrong — was not up to par. The season-to-date average entering Week 3 was 72% across 32 games. This week's performance was 22 percentage points below that baseline, and the numbers offer no soft interpretation.
Every upset this week was brutal. Eight of 16 games went against the favorite, and the model sided with the consensus favorite in all eight, agreeing perfectly with the market on every single one. When the market struggles this badly — also 50% — the week itself was hard to predict, not just for this model. But that distinction matters less when you're the one putting out a prediction.
The highest-confidence miss was SEA @ WAS. The model backed SEA at 74% confidence while giving an 8.5-point edge; WAS won 33-31, a 10.5-point swing. WAS entered missing its starting QB Jayden Daniels (elbow) and four other key contributors. SEA lost Julian Love and Ty Okada at safety. The game was decided by 2 points.
The biggest projection-vs-reality gap came not in a loss but a win: NE @ JAX. The model correctly called JAX but projected only a 3-point margin. JAX won 35-6 — a 26-point miss in the other direction. NE's roster was depleted entering the game, with three defensive players sidelined.
The model's average margin error for the week was 8.3 points. Six of eight losses came on games where the model's confidence ranged from 54% to 63% — low to moderate — suggesting some calibration problems at that tier, not just one catastrophic miss. The consistency of underestimating upset likelihood across the entire slate is the real story here.
This was not a week where sound reasoning met bad luck. This was simply a bad week.