The Sports PageWe Found a Pattern in Forty Games. Then We Killed It.
Penn State returned 7.6% of last year’s offense, hired a new coach, lost 48 players — and won 45–0. Clemson returned 41%, kept everybody, and lost by 41. The pattern held across the whole week. Then we gave it eleven seasons and it fell apart.
The question: does a club that keeps last year’s players beat its preseason rating?
The answer, briefly: no. And for one week it looked like a resounding yes.
Here is what is surprising. Not the result. The speed at which a very good story assembled itself out of nothing, and how nearly it got published.
What you get: the test that kills a pattern before it costs you money.
The story wrote itself
Open the week one box scores and a theory jumps out fully formed.
Penn State returned 7.6% of last season’s production. New head coach. Forty-eight players out of the door. By every measure of continuity they were a stranger to themselves, and they beat Marshall 45–0.
Clemson was the mirror. 41% of production back, the same staff, barely a transfer either way. They lost to LSU by 41.
Rutgers returned 44% and lost to Massachusetts, who returned a negative share — they lost more than they made last year. So we tested it properly across all 49 rated games: −0.276, and a t of nearly two. Clubs that kept more of last year played worse than the ratings expected.
That is a piece. It has a mechanism, two perfect anecdotes and a number attached.
Then we ran it ten more times
The same test, every opening week back to 2015. Same method, same ratings, same measure. Four hundred and ninety-five games.
Pooled, the correlation is −0.089. That is under one percent of the variation in how wrong the ratings were — a rounding error with a t-statistic.
Worse, look at where the dots sit. Four seasons lean one way; three lean the other; four sit on nothing. A relationship that changes sign every couple of years is not a small effect being measured imprecisely; it is the absence of an effect, seen forty games at a time.
Forty games is enough to find almost anything. It is nowhere near enough to believe it.
The ProfessorAnd there was a second way to be wrong
Suppose the effect had held. We would still have had the story backwards, and this is the part worth carrying.
The rating we tested against already contains returning production. It is one of the inputs used to build the preseason number.
So if clubs with lots of returning production underperform their rating, that does not mean the rating forgot something. It means the rating leaned too hard on something it already had. Those are different faults with different repairs, and the correlation cannot tell them apart.
We were hunting a missing ingredient. The same evidence fits a doubled one.
Why we went looking at all
Because the strategy is sound even when the first answer is not.
An extreme result is where a model announces what it does not know. When a rating misses by five touchdowns, something it does not contain just decided a football match. Collect enough of those and you can find out what.
So we are keeping a catalogue. Every game this season that misses its number badly gets logged with six possible explanations attached — including one nobody publishes: not how many transfers a club took, but how good the programmes were that they left. Twenty players out of a strong programme is not twenty players out of nowhere.
By November there will be six hundred games in it. Today there are fifty-one, and every line reads RUMOR on purpose.
What to take home: try to kill it before you print it
When a pattern appears, the instinct is to explain it. Ask a colder question first: where else should this be true?
Then go there and look. Not for confirmation; for the case that breaks it. We had a real number, a mechanism and two vivid examples, and it took one afternoon and ten more seasons to establish that we had nothing.
That is not a wasted afternoon; it is the cheapest thing we did all week. The expensive version is finding out in November, in public, after building three more pieces on top of it.
Every headline you read this weekend will be a pattern found in about forty games. Almost none of them were run eleven times.
Notes & sources
All figures from the CollegeFootballData API: game results, preseason SP+ ratings, returning-production shares (percent of prior-season PPA returning), transfer-portal moves and coach hire dates. Surprise is the actual margin minus the expected margin, where expected is the home club’s SP+ minus the away club’s plus 2.5 points for home field, dropped at neutral sites.
Week one 2026: 49 games where both clubs carried a rating and a returning-production figure. Correlation between the difference in returning production and the surprise: −0.276, t = −1.97. The named cases — Penn State 7.6% returning with a first-year head coach, 38 transfers in and 48 out, beating Marshall 45–0; Clemson 41.4% returning, no coaching change, 11 in and 15 out, losing 51–10 at LSU — are real and verified, and were chosen after the fact because they pointed the way we already suspected. That is the problem with them.
The eleven-season test covers every week one from 2015 to 2026, excluding 2020, on identical method: 495 games, pooled correlation −0.089, t = −1.98. Season by season: −0.376, −0.038, −0.303, +0.188, −0.271, −0.035, −0.099, −0.053, −0.000, +0.081, −0.276.
Two honest limits. The pooled t clears the conventional threshold, so a very small real effect cannot be ruled out — we are claiming it is far too small and far too unstable to explain anything, not that it is exactly zero. And week-one games are not independent within a season, since the same programmes recur; a properly clustered interval would be wider still, which only strengthens the conclusion.
The catalogue runs with python3 scripts/outlier_watch.py and reports the sample size beside every correlation, labelling anything under 150 games a rumor.
