When the Record Starts to Mean Something: A Raw Record Loses to a Coin Flip Until November

Two weeks from now the sport starts again, and within a month your feed will fill with 4–0 teams that have “announced themselves” and 1–3 teams whose coaches are done. So before the noise arrives we ran the question through the 877 completed 2024 games in this site’s cache: after N games, how well does a team’s record actually predict the rest of its season? The answer is rude. Used as-is, the record-so-far predicts rest-of-season winning worse than a flat guess of .500 until game 8 — November football — and the fix isn’t waiting, it’s shrinking: blend the record with four imaginary games of .500 and it beats knowing-nothing from game 3 on. Meanwhile the scoreboard is ahead of the standings the whole way: three games of scoring margins carry about as much rest-of-season signal as six games of wins and losses.

The test, stated honestly

For each of the 134 FBS teams (the house 8+-appearance proxy), line up its cached games in Eastern-time order. Take the first N as known, and try to predict its winning percentage over everything after N — requiring at least four remaining games so the target is a real sample. Score predictions by mean absolute error. Three predictors compete: the team’s raw win% through N; the number you’d use knowing nothing, a flat .500; and the shrunk record — the record plus four phantom games of .500, which is (W + 2) / (N + 4). N runs 1 through 8; beyond that the only teams with enough games left are the championship-and-playoff tail, a selection effect rather than a sample, so the analysis stops where the sample stays whole (n = 132–134 throughout).

Why a real record loses to a fake number

How can a team’s actual results predict its future worse than a number containing no information? Because a young record is an extreme number. After one game every team in America is 1.000 or 0.000, and no team on earth plays .000 football the rest of the way; the raw record’s error at N = 1 (0.466) is almost entirely self-inflicted overstatement. Rest-of-season win rates cluster toward the middle, so a predictor that lives at the edges pays for its confidence every single week. This is regression to the mean doing what it always does, and college football makes it worse than most sports: the early schedule is stuffed with buy games that inflate records without discriminating between teams — opening week alone was 61% FCS visitors in this cache.

The shrinkage estimator is nothing but that insight wearing arithmetic. Four phantom games of .500 is a crude prior — no recruiting composite, no returning production, no coach’s seat temperature — and it still beats both the raw record and the flat guess from game 3 onward. The lesson isn’t that this exact k is sacred; it’s that any respect paid to the middle beats treating a September record as a measurement.

4–0 means “pretty good,” not “great”

The same cache, cut the way fans actually talk. 23 teams started 2024 four-and-oh. Their combined record the rest of the way was .640 football — clearly good, nowhere near their undefeated advertising. The five 0–4 teams played .325 afterward; the thirty-one 2–2 teams, .524. So a perfect month bought about 12 points of future win rate over mediocrity, and a winless one cost about 20. Real signal, heavily discounted — and the tails supplied the object lessons. Utah started 4–0 and won 12% of its remaining games. New Mexico started 0–4 and played .620 the rest of the way — better than the average 4–0 starter. Neither team was lying in September; September itself just doesn’t say much.

The scoreboard leads the standings

The right panel is the finding we’d actually carry into this season. At every horizon from three games on, mean scoring margin through N predicts rest-of-season winning better than win% through N does — r = 0.38 vs 0.30 at three games, 0.45 vs 0.38 at five. Wins are a one-bit summary of a forty-point information stream; margins keep what the standings throw away, which is why Pythagorean expectation works and why gaudy one-score records regress. The cleanest 2024 illustration: Tulane sat 1–2 after three games — while outscoring opponents by 10 a game. The record said struggling; the margin said good team with hard losses. Tulane played .800 football the rest of the season. (Margins are not an oracle — they run through the same buy-game inflation as everything else in September — but they are consistently the better half of the scoreboard.)

One caveat belongs in the open: this is one season’s worth of teams, so every correlation above wears a standard error of about 0.09, and the exact crossover points would wobble in another year. What wouldn’t wobble is the ordering — extreme early records overstating themselves, shrinkage repairing them, margins leading wins — because those aren’t facts about 2024; they’re facts about small samples, and every September is one. The poll piece measured how long voters cling to August opinions; this is the other half of that trade — how long the results themselves deserve to be doubted.

What to do with this in September 2026

Three habits, all free. When a team starts 4–0, mentally book it as .640. When a record and its margins disagree, trust the margins — three weeks of them outrank six weeks of standings. And when you’re tempted to project anyone in September, remember the flat .500 guess out-predicted every raw record in America until November. The record starts meaning something around game three — but only after you’ve taught it some humility first.

How fast results beat the preseason poll

This section was first published on June 27, 2026 as a separate article.

A preseason poll is a guess. A well-informed guess — voters weigh returning starters, recruiting, coaching, last year’s film — but a guess made before a single snap. In the language of statistics it’s a prior: a belief you hold before the evidence arrives. Then the season starts, and every game is evidence. The question we care about is how fast that evidence should outweigh the guess, and there’s a clean answer. Bayesian updating says the season’s results overtake the preseason poll sooner than the polls themselves admit. Our position, stated plainly up front: voters anchor on the preseason ranking far too long, and the math says so.

One thing before the numbers, because the recent pieces on this site have all been honest about it: what follows is a clearly-labelled model, computed from a Bayesian updating model, not a database of poll movements. We’re showing you how a rational observer should reweight a prior as games pile up. That’s a normative claim, and we’ll be careful to keep it separate from what voters actually do.

The model in one line

Treat the preseason poll as worth k “virtual games” of information — as if, before the season, you’d already watched the team play k times. After n real games, your best estimate is a weighted average of the two sources, and the weights are just the counts. The weight on the season’s results is n / (k + n); the weight on the poll is k / (k + n). That’s the whole apparatus. The results outweigh the poll the moment n passes k, and the 50% crossover sits exactly at n = k.

The exhibit

Read the curves by trust level. If you give the poll moderate trust — worth k = 5 games — the season’s results hit 50% of the weight at n = 5 games, climb to 64% by nine games, and reach 71% by the end of a twelve-game regular season. Put another way, by season’s end the preseason poll is down to about 29% of your estimate. Trust the poll only lightly (k = 3) and the crossover comes at just three games. Even trust it heavily (k = 8) and the results still pull ahead: 53% at nine games, 60% at twelve. The headline we’d hang on this: under moderate trust, the season’s results outweigh the preseason poll after about five games — roughly early October — yet voters routinely keep top-five preseason teams parked near the top long after that.

Worked through, slowly

Take the moderate case, k = 5, and walk the season. After five games the weight on results is 5 / (5 + 5) = 0.50 — a dead tie between what you guessed and what you’ve seen. After nine games it’s 9 / (5 + 9) = 9/14 = 0.64. After twelve, 12 / (5 + 12) = 12/17 = 0.71. The arithmetic really is that plain; there’s no hidden machinery.

What does “the poll is worth five games” actually mean? It means the preseason guess is about as informative as having watched the team play five times. That’s the beautiful trick of treating a prior as virtual observations: it puts the guess and the games on the same scale, so combining them is just bookkeeping. Your posterior estimate is the weighted average (prior × k + results × n) / (k + n) — the prior counts for k games, the season counts for its n, and you divide by the total. Early on, when n is tiny, the prior dominates because it’s the biggest pile of “games” you have. As real games accumulate they swamp it. That’s not a quirk of football; it’s how conjugate priors behave everywhere.

Where this is too tidy

We won’t pretend the model is the territory. The honest caveats matter as much as the result.

• k is an assumption, not a measurement. How good preseason polls truly are is debatable, and a well-built preseason model — one leaning on returning production and recruiting — might genuinely be worth more than five games, especially in September when every team’s sample is microscopic. Set k too low and you’ll overreact to noise; too high and you’ll anchor like a poll voter. The choice of k is the whole argument.
• Not every game is equally informative. The model treats all n games as identical evidence, but a 49–3 win over a cupcake tells you almost nothing, while a road loss to a ranked team tells you plenty. Strength of schedule belongs inside n, and this version ignores it — see strength of schedule for why the opponent matters as much as the result.
• Results are noisy in the other direction. The opposite error to anchoring is overreacting — a 3–0 start against weak teams isn’t proof of a playoff team, just as a fluky 1–2 start isn’t proof of collapse. Early records mislead too, which is the same small-sample lesson behind why a few carries can’t pin down a rushing average. The model corrects anchoring; it doesn’t license credulity.
• This is how you should update, not how voters do. The curves are a yardstick for rational reweighting. The gap between them and the actual poll movement is precisely our complaint, not a description of it.

Where the idea comes from

The engine here is Bayesian updating, and specifically the conjugate-prior trick of expressing a prior as a number of virtual observations — the cleanest way to make “a prior worth k games” literal. It’s the same responsiveness dial that shows up in rating systems: a low k here is a fast-moving rating, and a high k is a sticky one, which is exactly the tension we unpacked in the K-factor and rating responsiveness. The empirical backstop is real, too: preseason rankings stay predictive surprisingly far into the season, largely because the things that drive them — returning production chief among them — are genuinely sticky year to year. That’s the honest case for a non-trivial k. It’s also why we land on “voters anchor too long” rather than “ignore the poll” — the prior earns its early weight; it just shouldn’t keep it.

Reproduce it

There’s nothing to hand-enter. The weight on results is weight = n / (k + n), plotted across a season for several values of k, and the figure is regenerated by charts/chart_prior_vs_data.py. Change k, change the season length, and the crossover moves to n = k every time. If you’d rather watch the same updating logic run inside an actual rating, the Elo and rating tools on the calculators page are the same machinery with the dial set for you.

The takeaway

A preseason poll deserves real weight in week one and almost none by November — and the handoff happens faster than the rankings act like it does. Under moderate trust the season takes over after about five games; even under heavy trust it wins out before the season’s halfway home. The poll was a fine prior. The games are the evidence. Let the evidence win on schedule.

Sources & further reading

• The weight curves are computed by charts/chart_prior_vs_data.py from weight = n / (k + n) for k = 3, 5, and 8 — a labelled Bayesian model, not game data.
• Background: the conjugate-prior idea of treating a prior as k virtual observations is standard Bayesian updating — any introductory Bayesian text (or the chapter above) covers the prior-as-pseudocount construction used here.
• Related: why returning production stays predictive, the K-factor and how fast a rating should move, and strength of schedule — the context the model leaves out.