The one analytics stat everyone quotes wrong on television

Every fall, a coach goes for it on fourth-and-2 from his own 38, the play fails, and within ninety seconds someone on television says the model was wrong. The model was not wrong. The model was never making a claim about that play.

Expected-value frameworks — win probability added, expected points added, expected goals, run expectancy — are all descriptions of a distribution. They tell you what happens on average across thousands of similar situations. A single outcome sitting in the tail of that distribution is not evidence against the average; it is a required feature of it.

Where the honest criticism lives is in the inputs. Generic fourth-down models assume a league-average offense against a league-average defense in neutral conditions. Feed them a backup quarterback, a 22 mph crosswind and an offensive line missing both tackles, and the recommendation should move. Good analytics departments adjust for exactly this. Broadcast graphics almost never do.

The second failure is time horizon. Win probability is indifferent to whether you are a contender maximizing this season or a rebuilding team where developing a young quarterback in high-leverage spots is worth more than the game itself. The number optimizes for one game. Front offices optimize for five years.

None of this is an argument against the math. It is an argument for using it the way the people who build it do: as a prior you start from, adjusted for the personnel actually on the field, and never as a verdict on one snap.