Win Probability and Leverage: Reading the Live Number

The Number Behind the Broadcast Graphic

The live win-probability (WP) number - the one that flashes on screen as "Home 72% to win" and lurches up or down after every big play - is now a fixture of NFL broadcasts and analytics sites. It looks like a verdict, but it is something more honest and more useful: a model estimate of each team's chance to win, given exactly where the game stands right now.

Understanding what that number is built from - and what it is not - turns it from a passive scoreboard ornament into a tool. It powers the way analysts assign credit to individual plays, it defines a high-leverage moment, and it comes with caveats that explain why two reputable models can show different numbers for the same snap.

The short version: WP is a trained estimate, not a fact. It reads the game state, compares it to thousands of historical games in similar spots, and reports how often the team in your position has gone on to win.

What Goes Into the Estimate

A win-probability model takes the current situation and returns each team's probability of winning. The inputs are the handful of things that actually determine how hard it is to win from here:

Score & time

The margin and how many seconds remain to erase or protect it

Down & distance

Whether this possession is likely to continue

Field position

How close the ball is to a score, plus timeouts and possession

The model is trained on tens of thousands of historical games. For any given state - say, down four at midfield, three minutes left, second-and-six - it effectively asks: of all the historical situations that looked like this, what fraction did the team in this position eventually win? That empirical answer, smoothed by a statistical model rather than read off raw buckets, is the win probability.

Some models add one more ingredient: team strength. A version that knows one side is a heavy favorite nudges its estimate accordingly; a "vanilla" version assumes two league-average teams and reads only the situation. Whether team quality is baked in is a main reason two models disagree, as we return to below.

WP(team) = P(this team wins | score, time, down, distance, field position [, team strength])

From WP Changes to Credit: Win Probability Added

The static number is interesting; the change in it is where the analysis lives. Every play moves the game from one state to another, and therefore from one win probability to another. The size of that move is Win Probability Added (WPA) - the swing a single play produced.

WPA(play) = WP(after the play) − WP(before the play)

WPA assigns credit (or blame) for each play by exactly how much it shifted the chance of winning. A routine first-down completion in the first quarter might add a fraction of a percent; a fourth-quarter touchdown that flips a deficit into a lead can add thirty or forty percentage points in one snap. Summed up, WPA gives a "who actually moved the needle" accounting that weights plays by their importance to the result, not by raw yardage.

WPA vs. EPA: expected points added (EPA) measures a play's value in points regardless of game state, while WPA measures its value in win chance, which depends heavily on score and time. A 20-yard gain has roughly constant EPA whether you trail by 3 or by 30 - but its WPA is enormous in the first case and nearly zero in the second. See EPA vs. DVOA for the points-based side of this.

Leverage: Not All Moments Are Equal

WPA naturally produces the idea of leverage - a measure of how much a given moment can move win probability. A play has high leverage when the range of plausible outcomes spans a wide swing in WP; it has low leverage when nothing that happens will change the win chance much.

A fourth-and-goal in a tie game with two minutes left is maximum leverage: the outcomes (touchdown vs. turnover on downs) sit far apart in win probability, so the play is decisive. A first-down handoff while up 28 in the fourth quarter is near-zero leverage: success and failure barely separate, because the game is decided either way. Leverage is why a "clutch" play feels different from a meaningless one - more of the outcome is literally at stake.

This connects directly to strategy. A fourth-down decision can be worth so much precisely because it often arrives at a high-leverage moment, where getting it right or wrong moves win probability a great deal - the machinery behind our fourth-down decisions explainer.

One line: WPA is the swing that did happen; leverage is the swing that could happen. High-leverage plays are where games are won and lost.

Why Two Models Disagree

If WP were a fact, every model would print the same number. They do not, and the reasons are instructive rather than alarming.

Different inputs and training

Models are trained on different ranges of seasons and use slightly different feature sets and smoothing. The same situation can map to modestly different probabilities depending on what history the model learned from and how it was fit.

Team quality: in or out?

A model that bakes in team strength will favor a strong team in a close situation; a situation-only model treats both sides as average. For a lopsided matchup in a tight spot, these two can differ by a wide margin - and neither is "wrong," they are answering slightly different questions.

Read a WP number as the output of a model with specific assumptions, not a universal truth. When two trackers disagree, the gap usually traces to whether team quality is included and what data the model learned from - both knowable, both reasonable choices.

A Late Turnover Swings the Number

The clearest way to feel how WP works is to watch a single high-leverage play move it. The figures below are invented for illustration, chosen to show a realistic shape rather than a real game.

Hypothetical situation, illustrative numbers only.

Setup: Team A leads by 4 with about two minutes left. Team B has driven into Team A's red zone, first-and-goal, and the win-probability model has the game close to a coin flip - Team B is one score from the lead, but time and downs are tight. Call it roughly:

~52%

Team B win probability, first-and-goal, before the snap

~6%

Team B win probability immediately after an interception in the end zone

The swing: the interception hands the ball back to the team that was already ahead and now only needs to run out the clock. In one play, Team B's win chance collapses from a near coin flip to a sliver, and Team A's rises by the mirror amount. The WPA of that single defensive play is on the order of 0.06 − 0.52 ≈ −0.46 for Team B - close to half a win on one snap. That enormous figure is possible only because the play happened at maximum leverage; the same interception in the first quarter would barely register.

This is the entire WP/WPA/leverage story in one play: the live number reads the state, the change in the number measures the play's impact, and the magnitude of that change reflects how much was at stake.

Honest Caveats

WP is powerful but easy to over-read. A few limits keep it in its lane:

Early-game WP is close to the pregame prior. In the first quarter little has happened, so the number barely moves off where the game started. A 55% reading on the opening drive mostly says "this team was a slight favorite before kickoff," not a fresh in-game judgment.
Most models assume average teams unless adjusted. A situation-only model does not know one side is far better; it reads the state as if two league-average teams were playing, which can understate a strong favorite's true chance.
It is calibrated on history, not this roster. The model reflects how the average team fared from a spot, not how this offense, defense, or kicker performs. Personnel and matchup live outside the basic inputs.
Probabilities are not promises. A team at 6% still wins sometimes - that is what 6% means. A surprising comeback does not prove the number was wrong; low-probability events happen.

Read WP as a well-calibrated estimate with stated assumptions, and it is one of the most useful numbers in football. Read it as destiny, and it will mislead you on exactly the games people remember most.

The bottom line

The live win-probability number is a model estimate - built from the current score, time remaining, down and distance, field position, and sometimes team strength - of each side's chance to win, trained on tens of thousands of historical games in similar spots. The change in that number from play to play is Win Probability Added, which assigns credit to each play by the swing it produced, while leverage describes how much a given moment can move WP at all. Two models disagree mainly because of different inputs and training data and whether they bake in team quality. And the honest caveats matter: early-game WP barely departs from the pregame prior, basic models assume average teams, and a probability is an estimate, not a guarantee. Used with those limits in mind, WP turns the scoreboard into a running measure of who is actually winning.