What the number answers
Two gains of the same length can have different value. Five yards that earn a first down are not the same situation as five yards that leave an offense facing fourth-and-long. Expected Points Added compares expected scoring value before and after a play, including scoring where relevant. Its usefulness comes from supplying that context, not from declaring raw yardage worthless.
EPA is itself a model output. The expected-points model, possession convention and treatment of penalties all matter. A figure copied from a dashboard without its definition is incomplete evidence. A positive average does not mean a team will score on its next possession or win its next game.
A deliberately hypothetical example
Imagine a non-scoring play starts with expected scoring value 0.9 and ends at 1.4 for the same offense. The change is +0.5. Those invented inputs illustrate subtraction; they are not an NFL lookup table, a measured league average or an output from our live engine.
Now consider an average from a handful of plays. One unusual turnover return can move that average sharply. Before comparing teams, check the number of observations, whether kneels and spikes are included, the opponents, and whether the source separates passing from rushing. Changing those filters can change the comparison without either calculation being arithmetically wrong.
How to use the information on Pick6Labs
Start with a team or game page, identify the season and source timestamp, and distinguish a source statistic from a forecast. The game engine uses team-efficiency inputs alongside other components; the presence of an EPA concept in this guide does not establish that every personnel-specific EPA split is available in production.
The Players Hub uses a separate position-relative box-score index. It does not measure EPA, blocking quality or coverage assignments. Check the rating evidence before comparing two players. A missing advanced statistic stays missing; an AI explanation should not be treated as a measurement that fills the gap.
An interpretation checklist
A useful note records the source, season, filters, sample size and the question being asked. For example: compare the same regular-season passing subset for both teams, then describe uncertainty rather than assigning an unsupported points adjustment. If any of those ingredients are unavailable, say so.
No historical claim that one metric is a fixed number of times more predictive than another is established here. That would require a defined dataset, an out-of-sample test and reproducible results. This guide explains how to read a metric; it is not a published backtest.