How to read these numbers

The numbers on a game

A game page carries a win likelihood, a projected margin, and a range. Worked through on an illustrative call of 62% for the home side:

What 62% means

Out of 100 games like this one, our model expects the home side to win about 62. It is a forecast, not advice.

Projected margin

Half the time we expect the home side to win by more than 4 points, half the time by less.

The range

Nine times in ten the final margin should land between the away side by 9 and the home side by 17. A wide range means the model is unsure.

How the forecast moved

Each point is a forecast we published this week. Earlier numbers stay on the page; we never overwrite a call.

Calibration, the claim we actually make

Calibration means the numbers mean what they say. When we publish 60%, that side should win about 60 times in every 100 such calls. When we publish 90%, about 90 in 100. A calibrated 62% is not a promise about one game; it is a promise about every game we have ever labelled 62%, checkable against the record.

That is a stronger claim than being usually right. Most NFL games are close, so a model can be right most of the time just by picking favorites, while its stated numbers drift far from reality. Calibration is the harder test: sort every published call into buckets by what we said, then ask whether each bucket landed where we said it would. Being right about how often we are right is the product.

A model that is often right and still worthless

Imagine a model that says 55% for the home side in every single game. Home teams do win a little more than half the time, so this model is right more often than wrong, and a scoreboard of its results looks respectable. It has also told you nothing: every game gets the same number, a coin-flip call and a mismatch read identically, and the one thing you wanted from a forecast, which games are different from the others, never appears. The uncomfortable part is that calibration cannot expose it either. Home sides win about 55 times in 100, so its one bucket lands exactly where it said it would: by the definition above, this model is perfectly calibrated. Calibration is necessary, and this model is the proof that it is not sufficient. What it lacks is sharpness, the credit a forecast earns for telling games apart, which a constant number never earns. So we grade with a score that prices both.

The Brier score, in plain language

The Brier score is how we grade forecasts. Turn the outcome into 1 or 0, take the gap between the forecast and the outcome, square it, and average across every call. Saying 90% on a game the favorite then loses costs far more than saying 60% on the same game; squaring makes confident misses expensive on purpose. Lower is better. A model that shrugs 50% at everything scores 0.25, so the score only rewards confidence that survives contact with results. The always-55% model lands at about 0.2475, barely ahead of the shrug: that sliver is its calibration counting for something, and the gap it never closes is the price of the sharpness it does not have.

How a published call is scored

Calls publish on a schedule, carry their publish stamp, and stay on the page after the result. The track record scores all of them, uncorrected.

Favorite won

The side we gave more than 55% won. Expected, though not certain.

Underdog won

The side we gave less than 45% won. When we say 42%, this should happen about 42 times in 100; it counts against us in the track record like every other call.

Coin flip

We gave neither side more than 55%, so no side was ours.

Published

When the forecasts on this page were computed and published. Forecasts change only on our schedule, not on every headline.

The numbers on a player

Player projections state a median and a range rather than one certain-looking value, and the data page documents the published percentiles.

This week's projection

The median of the outcomes our model simulated for this player this week. Half land above it, half below.

Points per game

Points scored and points allowed per game this season, counted from final scores only.

Injuries that matter

Players listed Out or Doubtful whom our rankings treat as starters for this team.

Why no DST or K rankings

We built a defence ranking, tested it, and it beat guessing: it picked the right top five about 26 percent of the time, against 16 percent for picking at random. Beating guessing was not enough to clear the bar we set for publishing it before we ran the test, so we do not publish one. We have not built a kicker ranking at all: it needs data we do not yet collect and a model that does not exist yet, and it will face the same bar when we do.

Check our work

The track record holds every published game call scored against its result. The data and methodology page names the inputs and the published artifacts. The changelog records what changed and when. If a number on this site cannot be traced to one of those three pages, treat that as our bug, not your misreading.