A statistical match centre

Football, taken apart

Down & Distance is the NFL sister to The Breakdown: results, rest, roof and the closing spread, treated as samples rather than recaps. The closing line wins; the ledger is the evidence. Elo, Massey, rolling PF/PA, Ridge, logistic and Poisson are extra lab scorelines and do not spend alpha.

What this is

Each finished game is a sample: a score, a closing spread and total, rest days, roof and surface. Collect those samples and a club or a season is a distribution, not a single Sunday.

Schedules come from nflverse — documented, versioned play-by-play and results from 1999, not a scraped vendor.

How to read it

Season is this year at a glance. Fixtures is every game, filterable. Structure is the league tree. Teams are cut by conference and division. Players is skill-position weeks. Years is the same file as a time window, with the 2020, 2021 and 2024–25 discontinuities labelled.

The spread names the favourite with a minus (Falcons −3); pick'em is PK. The file still stores the nflverse home-relative close. Blank lines on future weeks stay blank.

The close

C0 is the posted close, not a learned model. That is the thesis, not a placeholder. Scoreline challengers walked and did not beat it. Slot 4 is a wind shade of the posted total — the one place we looked for daylight — and it stays a challenger. Elo, Massey, rolling PF/PA, Ridge, logistic and Poisson are lab methods and are not ledger tests.

Players is weekly box scores. Expected points and residual use ffopportunity versus actual. Game Predictions publishes the close. Compare overlays clubs, divisions, conferences, seasons, or players.

Next up

Soonest remaining fixtures. The close appears once nflverse has posted it — usually a couple of weeks ahead, not on the season release.

Ways in

Every page is a different cut of the same schedule file.

Season This season at a glance Match counts, scoring by week, latest results and what is still scheduled.

Fixtures Every result Search and filter by team, season, week, status and game type, then open the close.

Teams One side, game by game Records, points and ATS against the close, grouped by conference and division.

Structure The league tree Conferences, divisions and all 32 clubs. Click a team for its record, scoring and ATS.

Years Cut by season 1999 onward, with era flags for the pandemic year, the 17-game expansion and the kickoff rewrite.

Players Box scores and fantasy Weekly PPR, expected points and residual. Display only — not a C1 or C2 feature.

Predictions The close, on the record Live 2026 scorelines from the posted line. Holdout lab and the spent α ledger sit beside it.

Compare Overlay clubs and seasons Points scored and allowed, scoring timelines, and weekly PPR. Filter by conference, division, team and player.

Season

This season

Live 2026 results and remaining fixtures. Switch to 2025 for the complete holdout year.

Season

Match status

Points per finished game, by week

Rest vs the close

Cover residual (result − spread) by home rest advantage. Short week is two or more days fewer rest than the visitor. Positive is a home cover. Gold dashed line is a push.

Scoring by roof

Combined points in finished games. Retractable includes open and closed lids.

Scoring by surface

Same totals, split grass vs turf. Small multiples, not a 32-club overlay.

Latest results

View all

Catalogue

All matches

Search by team. The close sits beside the scoreline; TBC until kickoff, and blank until nflverse posts the line.

Search

Status

Season

Week

Type

Explore

Teams

Open a side for its record, points and whether it covered the close. Historical abbreviations (OAK, SD, STL) resolve to the current franchise.

Season

League

Structure

The 32 clubs sit in two conferences and eight divisions. Click a team to open its page: record, points scored and allowed, and whether it covered the close.

Season

Catalogue

Players

Skill-position weeks for . PPR sorts the table. Residual is PPR points minus ffopportunity expected PPR points.

Search

Season

Week

Position

Team

Explore

Years

Browse by NFL season. 2020 is the pandemic year, 2021 adds the 17th game, 2024–25 are separate kickoff regimes.

Walk-forward lab

Predictions

The closing line wins. Switch to Holdout lab for the evidence: a reliability diagram, the Brier split, fourteen seasons instead of one, and the residual probe. Lab ratings and lab ML emit upcoming scorelines; they are not tests and do not spend alpha.

Calibration, refinement and the overall Brier are one-vs-rest means (three-class Brier divided by 3). The multiclass figures are that split times three. Uniform guess is 0.667.

What the methods are

Published and ledger

C0 close: The market’s closing spread and total. Scores come from those lines. Win chance uses the vig-removed moneyline when both sides are posted, otherwise the spread. The draw mass is a walk-forward empirical tie rate (, fit on seasons before the one being predicted) — down from about 3% under the old ±0.5 band. A σ that depends on the total is near-inert and is only a fallback. This is the live board. It is not a learned model.
C1 blind: Two gradient-boosted trees predict margin and total from pre-game differentials — rest, Elo, Massey, a two-timescale EWMA of points for and against, roof and era flags. They never see the betting line. Walked 2012–2025 week by week; worse Brier than C0 in the pooled window and in most seasons, so not published.
C1 given close: The same trees, allowed to use the closing spread and total as features. Slot 2, not the headline C1. Walked 2012–2025; did not beat C0.
C2 structural: Models how a team scores — touchdowns, extra points, two-point conversions, field goals and safeties — then adds those events into a scoreline. Walked 2012–2025; did not beat the close.
C3 wind total: The close, plus a walk-forward wind shade of the posted total on outdoor games. Margin and win chance stay at C0. Slot 4, judged on 2002–2025 — the coverage of wind, not the scoreline window. O/U on outdoor bets; did not clear. Its own card is below. A challenger, not a C0 amendment.

Lab only — not ledger tests

Elo, Massey, rolling PF/PA: Walk-forward ratings. Elo is a margin-weighted rating with home-field; Massey is least squares on past margins, frozen for the week; rolling PF/PA blends each side’s recent scoring and conceding.
Lab ML — Ridge, logistic, Poisson: Ridge is a linear model of margin and total on the same differentials as C1. Logistic predicts home / draw / away probabilities directly. Poisson is an attack–defence scoring model; win chances come from the two Poisson distributions, not the shared C0 normal. Switch to Holdout lab for the 2012–2025 argument, or stay on Upcoming and pick a Lab ML chip for 2026 scorelines.

Reliability diagram

Stated probability against how often that class happened, 2012–2025. The dashed line is perfect calibration. C1 is sharper and systematically overconfident.

Brier split

Three-class Brier. Calibration and refinement are rescaled to that Brier so they add. ECE is in probability points, not Brier units.

Platt-scaling the close’s vig-free moneyline returns a ≈ 1.05, b ≈ −0.035 — the market is already calibrated. C1 is sharper and systematically overconfident: it loses on ignorance, not on a fixable squash.

Next fixtures

Home win probability, ranked. Dashed lines at 30% and 70% mark high-certainty sides. Bars in grass are outside that band.

Confusion matrix

Home / draw / away from the close (or the walked challenger) vs what happened in 2025.

Precision and recall

Per outcome, then a macro average across the three classes.

Predicted vs actual margin

Predicted home margin against what happened in 2025.

Favourite probability bands

When the method gives a side 60–65%, how often did that side win? Draws count as misses.

Over/under residual

Observed total minus predicted total. Positive is over the model. Gold dashed line is zero; N(μ, σ) when n ≥ 8.

Player market

Player predictions

Residual is PPR points minus ffopportunity expected PPR points. None of this enters C1 or C2.

Position

Overlay

Compare

Choose what to overlay — seasons, clubs, divisions, conferences, or players. Teams, divisions, conferences, and players use a From–To span; seasons overlay chosen years of one slice. The first load starts with four series. Whiskers are ±1 sample SD of games in that season, not a confidence interval. Fantasy remains display only.

Compare

From

To

Conference

Division

Team

Position

Add

Comparison metrics

Points scored

Regular-season finished games. Bars are sample density; a dashed N(μ, σ) is drawn when there are at least eight games.

Points allowed

Same games, from the other sideline. Dashed curve is the normal fit when n ≥ 8.

Scored by week

Club games, or the group mean. Whiskers are ±1 sample SD when a week has two or more samples.

Allowed by week

Same week means from the other sideline.

Scored by season

Mean points scored. Whiskers are ±1 sample SD of games in that season, not a confidence interval of the mean.

Allowed by season

Mean points allowed. Same ±1 sample SD whiskers.

Weekly PPR

Each week is a sample across the span. Skill-position PPR only; not a C1 or C2 feature. Dashed N(μ, σ) when n ≥ 8.

PPR by season

Mean weekly PPR. Whiskers are ±1 sample SD of weeks in that season, not a confidence interval.

What this is

How to read it

The close

Next up

Ways in

Match status

Points per finished game, by week

Rest vs the close

Scoring by roof

Scoring by surface

Latest results

What the methods are

Published and ledger

Lab only — not ledger tests

Why the close wins

Reliability diagram

Brier split

Per season · 2012–2025

What the close does not already know

C3 · slot 4

Next fixtures

2012–2025 · scoreline

Candidate ledger

Confusion matrix

Precision and recall

Predicted vs actual margin

Favourite probability bands

Over/under residual

Expected vs actual (PPR)

By position

Comparison metrics

Points scored

Points allowed

Scored by week

Allowed by week

Scored by season

Allowed by season

Weekly PPR

PPR by season

Match details

Player