Win Probability and WPA Explained: What the Percentage Can and Cannot Tell You
A win probability graphic is an estimate of a team's chance of winning from a particular game situation. It is not the score, a promise about the ending or a grade for the next play call. Understanding that distinction makes the familiar rising and falling line much more useful.
This Football Nation education guide was researched October 20, 2026. Every numerical game state, probability and decision example below is hypothetical. They are teaching inputs chosen by FN, not outputs from a current NFL game or a model we have run. All examples simplify the final outcome to win or loss.
What goes into the estimate
Win probability, abbreviated WP, depends on the model and its inputs. The current nflfastR calculator documentation lists score difference, time remaining, down, distance, field position and remaining timeouts among its inputs. It also distinguishes a WP output from a separate output that incorporates the pregame spread.
Ben Baldwin's model explanation describes how game state features enter the estimate. Read that documentation before comparing charts: two providers need not use identical information or assumptions. An output informed by the pregame spread is not interchangeable with one that omits it, even when both columns say win probability.
FN's reading checklist is simple. Identify the team, game moment, provider and model version. Check whether the number is before or after the play, whether team strength is included, and how the provider treats a tie. Never assume a rounded television display contains every detail of the underlying estimate.
Example one: percentage points are not percent change
Suppose a hypothetical model moves the same team's estimated chance from 20 percent before a play to 40 percent afterward. The difference is 20 percentage points. Written on a zero to one scale, it is 0.40 minus 0.20, or plus 0.20.
Quantity | Calculation | Meaning |
WP before | 20% | Estimated chance before |
WP after | 40% | Estimated chance after |
WPA | 40% minus 20% | Plus 20 percentage points |
Relative increase | 20 / 20 | 100% relative increase |
Win probability added, or WPA, measures the change in modeled win probability associated with the play. The nflWAR research distinguishes that change from expected points added, which concerns scoring value. In this example, saying the chance increased by 20 percent is ambiguous: the absolute increase is 20 percentage points, while the relative increase is 100 percent.
Keep the team perspective consistent. If possession changes, do not subtract one team's preplay chance from the other team's postplay chance and label it WPA for the first team. Use the same team on both sides and follow the provider's timing and event conventions.
What does an 80 percent chance actually mean?
It does not mean the favorite must win this particular game. As a calibration illustration, imagine 100 separate situations each assigned an 80 percent chance by a model. Outcomes around 80 wins across that group would be consistent with that probability. One loss alone cannot establish that the model was wrong, just as one win cannot establish that it was accurate.
That illustration is not a validation test of any real provider. It assumes a suitable collection of observations and ignores sampling uncertainty for simplicity. Brill, Yurko and Wyner's simulation study specifically warns that football observations within a game share an outcome, complicating estimation and uncertainty. A large play count is not the same as an equally large collection of independent game outcomes.
Example two: the same yards can have different leverage
Consider two hypothetical five yard runs. We stipulate that the first changes Team A's modeled WP from 37 to 38 percent. The second, in a different score and clock situation, changes its modeled WP from 37 to 55 percent. These are invented model values, not claimed estimates for any actual down or yard line.
Hypothetical play | WP change | Rushing points |
Five yard run A | Plus 1 percentage point | 0.5 |
Five yard run B | Plus 18 percentage points | 0.5 |
The fantasy column assumes 0.1 point per rushing yard and no other scoring. It is identical in both examples, while the modeled game value differs. Reception bonuses do not affect these rushing examples. A league with additional scoring categories needs its own calculation.
FN analysis: this is why a memorable high leverage play and a large fantasy contribution are different claims. WP helps describe the importance of the situation. It does not award the runner extra fantasy points because the game was close, and it does not isolate his contribution from the blocking and defensive response.
A decision must include the failure branch
A team deciding whether to attempt a fourth down cannot compare only the attractive outcome with punting. The NFL's Decision Guide combines conversion likelihood with the game consequences of possible outcomes. That is a decision framework, not an officiating rule or a guarantee of success.
For a deliberately simplified hypothetical choice, assume a conversion attempt succeeds 60 percent of the time. Assume the team's win chance would be 50 percent after success and 10 percent after failure. Also assume the entire punt option has an estimated win chance of 28 percent, already averaged over its possible outcomes.
The attempt's weighted win chance is 0.60 times 0.50 plus 0.40 times 0.10. That equals 0.34, or 34 percent. Under these stipulated assumptions it leads the punt option by six percentage points. Comparing 50 percent directly with 28 percent would ignore the possibility and cost of failure.
The assumptions matter. Holding the two postattempt values fixed, the break even success probability is 45 percent: 0.45 times 0.50 plus 0.55 times 0.10 equals 0.28. Below that threshold, the punt option leads in this toy model. This is arithmetic sensitivity, not a universal fourth down recommendation.
The strongest counterargument
A clean formula can conceal uncertain inputs. An injury, a matchup or an incomplete model may change the conversion estimate or the values assigned to the next situations. FN would ask whether the recommendation survives a reasonable range of assumptions before describing the choice as obvious.
But uncertainty does not make the framework useless. It tells us to separate an estimated advantage from certainty. If the hypothetical attempt fails, that bad result does not by itself prove the original decision was bad. If it succeeds, the good result does not prove the assumed 60 percent conversion probability was correct.
How to use the chart without overreading it
For game analysis, identify which events changed the situation and check their sequence. Do not use the largest WPA play as automatic proof of the best player: teammates and opponents helped create the result, while opportunity to participate in decisive situations was uneven.
For fantasy, return to routes, targets, carries and the league's scoring rules. A receiver's team gaining win probability does not tell you whether that receiver received any of the work. Likewise, a quarterback helping protect a lead can be valuable to his NFL team without accumulating a large fantasy score.
Common questions and next checkpoint
Is WP a sportsbook quote? No. This guide discusses model estimates and provides no betting prices. Is WPA the same as EPA? No: one changes estimated winning chance, the other modeled scoring value. Does a comeback disprove an earlier high probability? Not on its own. Was the winning decision necessarily the best decision beforehand? Not necessarily.
At the next game you review, save the provider, team perspective, precise game moment and the two unrounded probabilities when available. Calculate their difference in percentage points, then explain the football event that changed them. Keep the decision's alternatives and uncertainty beside the result. That is more informative than treating a percentage as a prediction that cannot lose.
Continue learning
Sources and information cutoff
Sources checked October 20, 2026 at 12:04:29 PM CDT (2026-10-20T17:04:29.090Z). Sources explain model definitions and limitations. No current game probabilities, sportsbook quotes or player leaderboard values are reproduced. All worked numbers are explicitly hypothetical FN inputs and calculations.
Comments