Glossary

Every term, in plain language

Every term this site uses, in plain language. If something on a matchup page didn't make sense, it's explained here.

The Model

Analyst override

Elo only knows scores, so it can't see a starting quarterback getting hurt on Wednesday. An analyst override adds or subtracts rating points for one game, with the reason written next to it on the matchup page. It never hides inside the math — it's its own line in the calculation ledger.

Backtest

The site replays all 272 games of 2025 using only the rating each team carried before that kickoff — no game ever informs its own forecast. The result (64.0% straight-up, Brier 0.222) is the honest baseline the 2026 tracker is judged against.

See also: Brier score, Log loss, Calibration

Brier score

The standard report card for probabilistic forecasts: square the gap between the predicted probability and what happened (1 for a win, 0 for a loss), then average. Always guessing 50/50 scores 0.25, so anything under that means the model knows something. The 2025 backtest scores 0.224.

See also: Log loss, Calibration, Backtest

Calibration

A model can pick winners and still be dishonest about its certainty. Calibration checks the claim against reality: group games by what the model said, and compare to what happened. The accuracy page draws both bars side by side for every band.

Coin flip

The lowest confidence band on this site: the model's favorite wins less than 55% of the time in its own estimation. Read these predictions as 'no real edge either way' — the pick is a lean of a few rating points, not an opinion.

See also: Confidence label

Confidence label

Fixed bands on the favorite's probability: under 55% is a coin flip, under 65% a lean, under 75% confident, above that strong. These are this site's own thresholds, not an industry standard — and they're judged every week on the accuracy page, where a 'Strong' call is expected to win much more often than a 'Lean'.

See also: Coin flip, Lean, Confident, Strong

Confident

The third confidence band on this site: the favorite has a clear edge, 65–75% to win. Still expect the underdog to take about one in four of these — that's what the number means.

See also: Confidence label

Elo rating

A rating system borrowed from chess: every team carries one number, and after each game the winner takes points from the loser. Beating a strong team is worth more than beating a weak one. On this site every team started 2025 at 1500 and ratings were replayed through the full season, so a 2026 rating is a real summary of everything a team did last year.

Elo trail

The small line chart showing where a team's rating stood after every game of 2025. Rising into December means the team was genuinely peaking; a flat line near 1500 means the model never found much to say. Trails across the league share one vertical scale so they're honestly comparable.

Home field advantage

Playing at home is worth real points, so the home team's rating gets a bonus before the win probability is computed. This site adds 35 points, fitted to 2025 results — home teams went 142-123 that year, so the classic 55-point figure was overpricing the host. The bonus is zero at neutral sites.

K-factor

The speed limit of the rating system. A high K-factor makes ratings jumpy — one upset swings them a lot. A low one makes them sluggish. This site uses K = 26, which was not guessed: it was swept against all 272 games of the 2025 season and validated on games it wasn't fit to.

League average

Elo scales are arbitrary, so this site anchors them at 1500. Every team begins a season replay at 1500, offseason regression pulls ratings back toward it, and a team above 1500 is better than a typical NFL team. Roughly, 1600 is a contender and 1400 is a team picking early in the draft.

Lean

The second confidence band on this site: the favorite has a real but modest edge, winning 55–65% of the time by the model's math. Roughly a field-goal-to-touchdown favorite.

See also: Confidence label

Log loss

Like the Brier score, but harsher on overconfidence: saying 90% and losing hurts much more than saying 55% and losing. It's the objective the model's constants were tuned to — a setting only survives the sweep if it improves log loss on games it wasn't fit to.

Margin of victory multiplier

Winning by 21 says more than winning by 3, so the Elo update is scaled by ln(|margin| + 1). But there's a dampener: when a heavy favorite blows out a bad team, the multiplier shrinks, so good teams can't farm rating points off weak opponents.

Monte Carlo simulation

Instead of predicting one future, the forecast plays every remaining game thousands of times, letting each result fall according to its win probability and updating ratings as it goes. A playoff percentage is simply the share of those simulated seasons in which the team made the field.

Offseason regression

Between seasons, rosters churn and luck washes out, so last year's extremes don't fully carry over. Every 2026 starting rating is pulled 25% of the way back toward 1500 before the blend is finalized. A team that finished 2025 at 1650 starts the new year at 1612.5, not 1650.

Playoff odds

The share of simulated seasons in which a team finishes as a division winner or one of the three wild cards in its conference. Because seeding depends on tiebreakers, the simulation sorts full standings on division record and then conference record before counting.

Point spread

The rating difference divided by 25, rounded to the nearest half point — so a 75-point edge shows up as '−3'. It puts the prediction in a language football fans already read. It is the model's own number, computed before it ever sees a sportsbook's line.

Projected record

The average of a team's wins across all simulations, shown with a 10th-to-90th percentile range. It is a centre of gravity, not a prediction of an exact record — a 9.4-win projection usually means anywhere from 7 to 12 wins is entirely normal.

Pythagorean expectation

A team that outscores opponents by 200 points is usually better than its record, and one that wins six one-score games is usually worse. Pythagorean expectation converts points for and against (PF^2.37 / (PF^2.37 + PA^2.37)) into an expected winning percentage, then into a rating. It's one of the measured inputs in each team's 2026 starting rating.

Rating derivation

The section on each matchup and team page that shows its work: the Elo trail through 2025, the blend of measured inputs, the 25% offseason regression, and any roster adjustment — ending in the exact starting rating used for predictions. Nothing in the number is hidden.

Rest advantage

A team coming off a bye or a Thursday game has more recovery time than one on a short week, and the model prices that gap at a few points per day, capped so it never dominates the matchup. Week 1 has no rest edge at all — every team is coming off the same offseason.

Roster adjustment

A 2026 judgment applied on top of the 2025 measurements: a retired quarterback, a blockbuster trade, anything Elo can't see because it happened after the season ended. It's capped at ±40 points, applied after offseason regression, and always displayed with its reason on the team page.

Strength of remaining schedule

The mean Elo rating of every opponent still on a team's schedule. Above 1500 means a harder run-in than average; below 1500 means an easier one. Two teams with the same record can have very different playoff odds purely because of this number.

Strength of schedule

The average rating of the opponents a team actually faced, ranked across the league. It keeps records honest: 11-6 against the league's hardest slate means more than 11-6 against its easiest. On matchup pages it also splits each team's record into games against above- and below-average opponents.

Strong

The highest confidence band on this site: the favorite wins more than 75% of the time in the model's estimation. Even here the underdog wins about one game in five, and the accuracy page tracks whether the band earns its name.

See also: Confidence label

Win probability

The headline number on every matchup. It comes from one logistic curve: take the rating gap (home rating plus home field, rest and any override, minus away rating), and P(home) = 1 / (1 + 10^(−Δ / 400)). A 100-point edge works out to about 64%; 200 points is about 76%.

Football

Bye week

Each team skips one week between Weeks 5 and 14. Coming out of a bye a team has roughly twice the normal rest, which the model prices as a rest advantage the following week.

Conference

The NFL splits its 32 teams into the American Football Conference and the National Football Conference. Teams play most of their schedule inside their own conference, and the conference champions meet in the Super Bowl.

Division

Each conference has four divisions (North, South, East, West) of four teams. Division rivals play each other twice a year, which is why the same matchup shows up every season and division games carry playoff weight.

Draft order

The 18 teams that miss the playoffs are ordered worst record first, so the team most likely to hold the #1 pick is the one the model expects to lose the most. The forecast reports how often each team lands the first pick and how often it lands in the top five.

Game log

On each team page, the full 2025 schedule with scores, opponent ratings, and the win probability the model would have posted before each kickoff — so you can see not just what happened, but what the model expected to happen.

Neutral site

Some games — international games, the Super Bowl — are played somewhere neither team calls home. The model detects these and sets the home-field advantage to zero, which is why a 'home' team at a neutral venue gets a slightly lower win probability.

Point differential

The simplest honest measure of a team: over 17 games, did it outscore its opponents? Point differential predicts future wins better than win-loss record does, which is why it feeds the scoring-strength term in the rating blend.

Regular season

Each team plays 17 games across 18 weeks, with one bye. This site covers Week 1 of the 2026 regular season; the ratings behind it were built from the full 2025 regular season.

Season type

The data source tags every game as preseason (type 1), regular season (type 2), or postseason (type 3). Only regular-season games feed the ratings on this site — preseason scores tell you nothing about team strength.

Seed

Playoff teams are seeded 1 through 7 in each conference by record, with the top seed earning a first-round bye. On this site the term mostly appears in the data pipeline, where postseason seeding is kept separate from the regular-season ratings.

Seed

Division winners take seeds 1 to 4 by record, the three best remaining teams take seeds 5 to 7. The #1 seed gets the only first-round bye and hosts every game it plays until the Super Bowl, which is why it moves title odds so much.

Week

The NFL schedules its regular season in 18 numbered weeks. This site is v1: it covers Week 1 of 2026 in full, with the 2025 season's weeks preserved in every team's game log.

Betting & Odds

Closing line

The market's last word — the opening line plus a week of betting and injury news. Beating the closing line consistently is the gold standard for a prediction model, and it's why this site never lets its numbers see the market's.

Moneyline

The betting-market version of a straight-up pick: each team gets odds reflecting its implied chance to win. The model's win probability is the same idea without a sportsbook's margin baked in. Shown for comparison only — not betting advice.

Opening line

The market's first opinion, posted days before kickoff. It moves as money and news arrive, which is why it's distinguished from the closing line. Where a market line appears on this site, it's for comparison only.

Over/under

A line on combined scoring: bettors take the over or the under. This model only predicts winners and margins, so the total appears on matchup pages purely as market context the model never sees.

Spread cover

A favorite 'covers' when it wins by more than the spread; an underdog covers when it loses by less or wins outright. The accuracy page tracks the model's spread error — how far its line landed from the final margin — which is the model-side version of the same idea.

Straight-up

The simplest way to grade a prediction: did the team the model favored actually win? The accuracy page leads with straight-up percentage because it's the number anyone can check. Nothing on this site is betting advice.

Vegas line

The consensus market spread, shown on matchup pages strictly as a comparison. The model's own spread is computed from ratings alone and never sees this number — the footer of every matchup says so.

Data & Site

Backfill

When a score, line, or venue is missing, the site re-fetches that week from ESPN's scoreboard API and fills the gap. It's how the 2025 season was loaded into the database and how late-arriving final scores get picked up.

Kickoff window

Games in a week cluster into familiar slots: Thursday night, Sunday 1:00 ET, Sunday 4:05/4:25 ET, Sunday night, and Monday night. The schedule page groups games by window (using Eastern time, the broadcast standard), while each game also shows the kickoff in your local time.

Ledger

On every matchup page, the calculation ledger lists each term in the prediction — home rating, away rating, home field, rest, any analyst override — and what it contributed, ending in the exact win probability shown up top. If a number on this site ever looks wrong, the ledger is where you audit it.

Pregame

The honesty rule behind every number on this site: a pregame win probability uses only the ratings as they stood before that kickoff. In the 2025 game logs, each probability is what the model would have posted that morning — never revised with the benefit of the result.

Slate

Broadcast shorthand for a week's complete schedule. The schedule page groups the slate by kickoff window — Thursday, Sunday early, Sunday late, Sunday night, Monday night — the way a broadcast desk would lay it out.