Our Model Is Sharpest in September — and Week 18 Breaks It
Ask any football fan when the NFL is hardest to predict and you'll get the same answer: September. New schemes, new rosters, no current-season data — surely the early weeks are a fog, and the picture sharpens as the season goes on.
Our model's graded record says it's exactly backwards.
We grade every prediction our machine-learning model makes against the closing spread and store the results — every week, every season. Break that lifetime record out by week of the season, and a pattern emerges that took us a while to accept:
| Stretch of season | Record | Accuracy |
|---|---|---|
| Weeks 1–3 | 158–121 | 56.6% |
| Weeks 4–13 | 454–391 | 53.7% |
| Weeks 14–18 | 202–215 | 48.4% |
| Weeks 16–18 | 110–127 | 46.4% |
| Week 18 alone | 20–27 | 42.7% |
The model's best stretch is the one everyone calls unreadable. Its worst — by a wide margin — is the stretch when we supposedly know the most, capped by Week 18, where it performs materially worse than a coin flip. The 2025 season told the same story in miniature: a 13–1 Week 5 and a 21–10 open through Weeks 2–3, followed by a 32–45 slide across Weeks 14–18.
Two questions, then: why does a statistical model peak exactly when human confidence bottoms out — and collapse exactly when human confidence peaks?
September: everyone else is looking at the wrong numbers
The September fog is real — but it only affects certain kinds of information. What's genuinely unreadable in September is the stuff of headlines: win-loss records, statement wins, "is this team for real?" narratives. Our own feature research quantified this: a team's early-season record is mostly noise until four or five games in — noisy enough that inside our feature pipeline we deliberately blend early-season team stats with prior expectations rather than take a 2-0 start at face value.
But the model doesn't watch headlines. Its features are rolling, multi-season efficiency measurements — passing and rushing efficiency differentials, form windows stretching back through last season — and those carry over the offseason far better than reputations do. Meanwhile, September expectations set by humans are anchored to last season's story: last year's playoff team opens as this year's default contender, the breakout that already happened gets priced like it will happen again.
That mismatch is the model's edge, and it's largest precisely when current-season evidence is scarcest. Early in the year, the model is quietly comparing its stable efficiency read against a consensus still running on last season's narrative. By mid-season everyone has seen ten games of evidence, the consensus has caught up, and the model's advantage compresses — from 56.6% to a still-solid 53.7%.
December and January: the games stop being about talent
Then the late season arrives, and the model doesn't just lose its advantage — it goes underwater. What changes isn't the math. It's the games.
By Weeks 16–18, a growing share of the slate is decided by things no efficiency statistic can see:
- Motivation asymmetry. Eliminated teams evaluating rookies, locked-in playoff seeds coasting, fringe teams playing desperate football. The talent gap the model measures becomes conditional on effort the model can't measure.
- Week 18 is a different sport. Starters rest, backups audition, coaches make decisions with one eye on the playoff bracket. Humans reading beat reports Friday afternoon know who's sitting; a statistical model trained on "how good are these teams" is answering a question the game is no longer asking. The 42.7% isn't a bug — it's a measurement of how little Week 18 resembles the football the model learned.
- Information symmetry. Whatever the model knows in December, the market consensus knows too — seventeen weeks of shared evidence closes the gap that September narratives left open.
We could hide this by quietly not publishing late-season predictions. Instead it's in the graded record on the site, because the shape of the failure is itself the insight.
What this means if you follow the NFL by the numbers
- Trust efficiency in September; distrust standings. Early-season records and narratives are the least reliable numbers of the year — but early-season efficiency against a consensus still anchored to last season is the most exploitable mismatch on the calendar. The season's clearest statistical reads come earlier than instinct says.
- In late December, context beats computation. Who's eliminated, who's resting, who needs the game — from Week 16 on, that context outweighs any power rating, including ours.
- Judge any prediction system by its full calendar. A model that only shows you its September should be asked about its January.
The honest caveats
- Per-week samples are ~85–95 graded predictions; individual week-to-week wiggles (a soft Week 8, a strong Week 13) are noise. The robust finding is the arc: a 10-point accuracy gap between the season's first three weeks and its last three, in a consistent direction, across the model's entire graded history.
- The early-season advantage is an argument about information lag, and information lags close. If the consensus starts weighting efficiency carryover properly in September, this pattern will shrink — one more reason we publish it rather than hoard it.
See the week-by-week record
The full accuracy tracker — every graded prediction by season and week, hot streaks and Week 18s alike — is on the SpreadTrends Predictions page.
Data: all graded model predictions in the SpreadTrends database (1,500+ across six seasons), scored against closing spreads. Pushes excluded from percentages. Playoff weeks omitted from stretch aggregates (n too small).