NFL Success Rate vs. EPA: Consistency or Explosive Efficiency?

Expected Points Added (EPA) measures the change in a team’s expected scoring value from before a play to after it, including the immediate value of any score. This is what EPA means in NFL analysis:

EPA = post-play expected points − pre-play expected points

The expectation model accounts for context such as down, distance, field position, time remaining, and score differential. A five-yard gain on third-and-4 can be valuable because it earns a first down; the same gain on third-and-12 may reduce expected points.

Unlike a binary success label, EPA intentionally preserves magnitude. A routine conversion might add 0.6 EPA, while an 80-yard touchdown could add more than 6. That explosive gain receives full credit rather than counting as merely one successful play.

Negative outcomes scale similarly:

• Sacks lose yardage and worsen down-and-distance, often producing sharply negative EPA.
• Turnovers transfer possession and may create favorable field position for the opponent, making them especially costly.
• Failed plays near scoring range can erase more value than comparable failures elsewhere.

EPA per play (EPA/play) is simply total EPA divided by plays. It estimates average efficiency across a sample while allowing rare, game-changing events to exert proportionate influence.

Success rate asks how often an offense clears a defined bar. EPA/play asks how much scoring value its plays create or lose on average. The difference is not which plays enter the calculation, but how each metric represents their outcomes.

A 12-yard completion and a 70-yard touchdown can both count as one successful play in success rate. The touchdown is not excluded; its extra value is simply compressed once the binary threshold has been crossed. Likewise, a failed play remains one failure whether it falls inches short or ends in a turnover, unless the chosen success-rate definition adds separate treatment.

EPA does not ignore routine gains. A four-yard run, short completion, or incompletion changes expected points according to down, distance, field position, time, and game state. EPA records that change even when it is small, then averages those values across plays.

That produces two distinct statistical views:

• Success rate emphasizes the share of plays meeting the criterion and is relatively insensitive to differences within the success and failure bins.
• EPA/play preserves the entire outcome distribution, making explosive gains, sacks, turnovers, and ordinary plays matter in proportion to their estimated scoring impact.

Thus, success rate describes consistency around a cutoff; EPA/play captures average efficiency, including the tails.

Assume two offenses each run five plays from broadly comparable situations. The EPA figures below are illustrative, but they show how identical average efficiency can conceal opposite performance profiles.

Both series produce +0.18 EPA per play. Under a positive-EPA definition, however, the steady offense posts a 100% success rate, while the volatile offense succeeds on only 20% of plays. One profile advances scoring expectation repeatedly; the other depends entirely on a single explosive.

Binary labels become even less descriptive when negative outcomes vary sharply. An incompletion worth roughly −0.3 EPA, a third-down sack worth −1.5 EPA, and a pick-six worth perhaps −6 EPA can all count as one failed play. Success rate treats them equally; EPA records how much damage each caused.

That distinction matters diagnostically. A low-success, high-EPA offense may be explosive but fragile. A high-success, modest-EPA offense may stay on schedule yet lack chunk gains. Neither metric resolves both questions alone.

Success rate and EPA/play form a useful diagnostic matrix. Neither axis supplies a complete verdict; each combination points toward different play-level evidence.

Disagreement between the metrics is therefore evidence, not noise. The next step is to inspect the play distribution: separate early downs from late downs, passing from rushing, routine failures from catastrophic ones, and open-field production from red-zone performance. That reveals whether the gap reflects a sustainable identity, situational weakness, or a small number of extreme plays.

Team-level averages blend fundamentally different situations. Passing usually produces higher EPA but greater variance than rushing, so an offense’s overall profile partly reflects its pass rate. Down-specific splits matter too: second-and-short invites aggression, while third-and-long creates sacks, checkdowns, and low success rates.

Useful diagnostic splits include:

• Red zone: Compressed space suppresses explosive gains, making success rate and finishing ability more informative than raw EPA/play alone.
• Neutral situations: Early-down plays before obvious clock or score pressure better isolate baseline offensive intent. Definitions vary, so score margin, quarter, and win-probability filters should be stated.
• Score state: Trailing teams pass into defenses protecting the sideline and deep field; leading teams run against loaded boxes. These selection effects can distort both metrics without indicating a true change in talent.

Pace requires separate treatment. More snaps primarily create opportunity, raising totals such as points, yards, and cumulative EPA; they do not mechanically improve EPA/play or success rate. Still, pace can indirectly shape efficiency through defensive fatigue, no-huddle communication stress, field position, and the extra opponent possessions that change game script.

The cleanest comparison therefore reports pass/run, down, field zone, neutral-state, and score-state splits alongside overall figures. Pace-adjusted rates should accompany counting production, not replace it.

Neither metric becomes trustworthy merely because it is expressed to three decimals. In small samples, one pick-six, fourth-down stop, or 70-yard touchdown can swing EPA/play; its distribution is heavily skewed, so the mean converges slowly. Splits by down, personnel, or quarter shrink the sample further and should carry uncertainty intervals.

Success rate is usually steadier, not immune. A play barely above a positive-EPA cutoff counts exactly like a major gain, while a tiny model revision can move borderline plays across the line. Yardage-based definitions introduce different threshold effects.

Provider choices also matter: expected-points models, garbage-time rules, kneel-downs, spikes, scrambles, penalties, and play attribution can all differ. Consequently, values from separate sites are not interchangeable, even when labels match.

Opponent quality belongs in the interpretation. Raw efficiency against elite defenses is not equivalent to the same mark against a weak slate, although strength-of-schedule adjustments can themselves be unstable early and depend on the adjustment method.

Finally, these are team-play outcomes, not context-free player grades. Line play, scheme, field position, receivers, quarterback decisions, and game state are entangled. Neither metric isolates individual talent without a richer model.