A win probability plot compresses an entire football game into a single line, usually one team’s estimated win probability tracked from kickoff to the final whistle. It looks simple to read — the line goes up, that team’s chances improved; the line goes down, they got worse — but a few features of the chart are easy to misread if the mechanics behind them aren’t clear.
The Two Axes
The horizontal axis is time, almost always running from kickoff to the final whistle, frequently marked with quarter boundaries so a jump can be located in context — a swing in the two-minute drill of the second quarter reads very differently from the same size swing at the start of the third. The vertical axis is win probability on a 0-100% scale, typically for one designated team (often the home team by convention), with the 50% mark as the natural midpoint: above it, the model currently favors that team; below it, the model favors the opponent. A line crossing 50% marks a change in which team the model currently favors, which is not the same thing as a change in who is actually ahead on the scoreboard — a team can trail 50%+ win probability while still leading the game, if the model expects that lead to erode.
nflfastR’s play-by-play data carries the columns that produce this chart directly: a wp column giving win probability for the possession team on that play, and a home_wp column giving it specifically for the home team regardless of possession — the same pair of numbers, computed play by play, that any win probability plot is simply drawing as a line (Open Source Football, cited below).
Why the Line Doesn’t Start at 50%
A win probability plot built on a spread-adjusted model, as described in Part 1 of the win-probability building series, doesn’t start at an uninformed 50-50 guess — it starts from the pregame point spread, since that public number already reflects team strength, injuries and home-field advantage before a single play has been run. A team favored by a wide margin shows a plot that opens well above 50% at kickoff, and that starting point is expected to fade toward whatever the in-game state supports as the model’s own decaying weight on the pregame prior shrinks with the clock.
A plot built on a model that skips the pregame prior entirely will instead open at exactly 50%, regardless of how lopsided the actual matchup is, and only start moving once the first few plays of the game produce a score, clock and field-position reading for the model to work with. Comparing two plots side by side — one spread-aware, one not — is one of the more visible ways to see what including a pregame prior actually buys a model: a flatter, more accurate first quarter instead of an artificially even one.
One Line Describes Both Teams at Once
A win probability plot for the home team and one for the away team are mirror images of each other rather than two independent pieces of information — at any point in the game, home win probability plus away win probability sums to 100%, since exactly one team wins. That means a single line, plotted for either team, already contains everything the chart has to say about both; a second line for the opponent would just be the same curve flipped around the 50% mark. Some charts plot both lines anyway, shaded to fill the gap between them, mainly for visual clarity rather than because the second line adds new information — the shaded gap makes it easier to see at a glance how far from even the game currently sits, without requiring the reader to do the 100-minus-the-number subtraction themselves.
Reading Swings Without Over-Reading Them
The size of a jump in a win probability plot depends on two things at once: what happened on the play, and when it happened. A fumble returned for a touchdown late in a one-score fourth-quarter game produces one of the largest single-play swings a plot will show, because it directly and immediately changes the outcome the model is trying to estimate. The identical play in the first quarter of a game that is already 21 points one way barely moves the line, because time remaining and the existing score cushion mean that single event has far less bearing on the eventual outcome — the same score-time interaction that shows up throughout the model-building series.
That dependence on timing is a feature of a well-built model, not a flaw in the chart. A model whose plot barely reacted to a late, genuinely decisive turnover would be under-reacting, not appropriately stable — the swing itself isn’t evidence of an unstable model. What would be worth questioning is a swing that seems out of proportion to how much a specific play should actually change a team’s chances; that is a modeling question about whether the feature set and calibration (covered in Part 3 of the building series) are capturing the situation correctly, not something the chart itself can diagnose.
A Single-Game Plot Is Not a Calibration Plot
It is easy to confuse a win probability plot — one line, tracking one game, over time — with a calibration or reliability diagram, which is a completely different chart built from many games at once, comparing predicted probability buckets against observed win rates, as described in the piece on what makes a win probability model good. A single-game plot cannot show whether a model is well calibrated, because calibration is a property of a model’s behavior across many comparable situations, not of any one game’s trajectory. A single dramatic swing or a single surprising finish on one game’s plot says nothing on its own about whether the model that produced it is trustworthy — that question can only be answered by the pooled, many-game version of the chart, not by staring harder at one line.
What a Plot Cannot Tell You
A win probability plot shows a point estimate over time, not the uncertainty behind each point — the same distinction covered in the piece on win probability, uncertainty and overfitting. A rare, unusual game state late in a close game can produce a confident-looking line even when the model has relatively few historical comparisons to base that specific estimate on. Reading a plot well means treating each point as the model’s best single estimate given everything known at that moment, not as a claim about how the game is guaranteed to end — the same caution that applies to any single probability number, whether it comes from a chart, a broadcast graphic, or a live win probability calculator. For a step-by-step walkthrough of how one of these charts gets built from a single game’s plays, see the explainers section.
Frequently asked questions
What does the x-axis of a win probability plot represent?
Time elapsed in the game, usually running left to right from kickoff to the final whistle, often marked with quarter boundaries. Some versions plot game seconds remaining instead, running the other direction, so it's worth checking which convention a given chart uses before reading it.
What does the y-axis of a win probability plot represent?
The estimated win probability for one team, typically the home team by convention, on a 0-100% scale. A value above 50% means the model currently favors that team; a value below 50% means it favors the opponent, and the line crossing the 50% mark marks a lead change in the model's estimate, not necessarily in the actual score.
Why does a win probability plot start away from 50% before the game even begins?
Because a model that uses a spread-adjusted pregame prior, as described in Part 1 of the win-probability building series, starts from the pregame point spread rather than from an uninformed 50-50 guess. A team favored by a wide margin will show a plot starting well above 50% at kickoff, before a single play has been run.
Do sharp vertical jumps in a win probability plot always mean the same thing?
No — a jump's size depends on both the play's outcome and when it happened. A turnover in the fourth quarter of a close game produces a much sharper jump than the same turnover in the first quarter of a game already decided, because time remaining amplifies how much any single event can swing the estimate, the same interaction described in the building series.
Is a big swing in a win probability plot proof the model is unstable?
Not by itself. A well-calibrated model is supposed to swing sharply on genuinely decisive plays — a pick-six with two minutes left should move the estimate a lot, because it directly changes the game's likely outcome. A swing only signals a problem if it happens on a play that shouldn't be that consequential, which is a modeling question, not a charting one.