The EPA Mirage: When a Team's Record Is Lying to You

The 2025 Kansas City Chiefs finished 6-11. They were also ahead of their opponents on all four of our EPA splits — offensive and defensive, pass and rush — for a net EPA of +0.121 per play, which across the previous 23 seasons has typically belonged to a 10-win team.

The 2025 San Francisco 49ers finished 12-5 with a net EPA of −0.026 — fractionally below zero, the efficiency profile of an 8-win team.

One of those records is describing the team that played. The other is describing a season's worth of coin flips. The interesting question is which one next September believes — and we can answer it, because we have 736 team-seasons where we know what happened next.

What the gap actually measures

Expected points added asks a narrow question on every snap: given down, distance and field position, how much did this play change the offense's expected points? Sum it up and you get a per-play efficiency number that ignores the two things a win-loss record is most distorted by — turnover luck and what happens in one-score games.

Our EPA ratings split four ways: offensive and defensive, pass and rush. Defensive figures are EPA allowed, so lower is better, and the net numbers are simply offense minus defense.

Across 2002–2025, net EPA and win percentage move together almost lockstep in the same season: r = +0.845.

What r means here. It measures how tightly two numbers move together, from −1 to +1, where 0 is no relationship and the sign gives the direction. The useful thing is knowing what counts as big in football, because almost nothing in this sport correlates strongly with anything:

• 0.85 — two ways of describing the same thing
• 0.33 — a real tendency you could never rely on for a single season
• 0.13 — visible across hundreds of team-seasons, invisible in any one of them

Every figure in this article sits somewhere on that ladder, and where it sits matters more than its sign.

So +0.845 is not a discovery — it is a sanity check. Efficient teams win. What matters is the residue: the distance between the record a team posted and the record its efficiency implies.

Does the gap predict anything?

Here is where the honest version diverges from the fun version.

Head-to-head, efficiency is not meaningfully better than record at forecasting next season:

That is a rounding difference across 736 pairs, and anyone who tells you efficiency dramatically outpredicts record is overselling it. Both are weak. Football is 17 games.

The signal shows up when you ask a better question: what does each number know that the other doesn't?

Efficiency carries information the record doesn't, and slightly more of it than the record carries beyond efficiency. That is the whole case for the mirage — and it is a tilt, not a revelation.

The regression pattern

Sort every team-season by how far its record ran ahead of or behind its efficiency, then look at what happened the following year:

Monotonic across all five buckets, and the correlation between the gap and the following year's change in win percentage is −0.399.

That table is seductive and it overstates the case, so here is the objection before someone else raises it: teams with .355 records improve and teams with .635 records decline anyway, efficiency or no efficiency. Records regress to the mean on their own. The table above cannot separate the two effects.

The test that isolates it

So hold the record roughly constant and vary only the efficiency underneath it.

Teams that won 10+ games, split by whether their efficiency backed the record:

Both groups land on exactly .573. The group whose record outran its efficiency simply started higher and had further to fall. A 12-win season built on efficiency and a 12-win season built on close-game results were worth the same thing the following year. Not similar — the same number, to three decimals, across 254 team-seasons.

Teams that won 6 or fewer, same split:

The unlucky group rebounds about five percentage points harder despite having started marginally worse.

Five points of win percentage is under one win a season. That is the true size of this effect, and it is worth knowing precisely because it is small enough to be missed and real enough to matter over several seasons.

2025's biggest gaps

Expected wins below are what each team's net EPA implies over a 17-game season, using the 2002–2025 fit.

Records ahead of their efficiency:

Efficiency ahead of their record:

The distinction that matters

Those two lists look symmetrical. They are not, and this is where most "regression candidate" lists go wrong.

A team can under-perform its efficiency two very different ways:

• The Chiefs and the Colts posted positive net EPA — genuinely good process, results that didn't follow. Kansas City's +0.121 came with a passing game generating +0.039 offensive EPA per play against a defense allowing −0.002; that is a good team that lost eleven games.
• The Cardinals and the Giants posted negative net EPA and finished below even that. Arizona's defense allowed +0.158 EPA per pass play — one of the worst marks in the league. They were bad, and then things went worse than bad. Their rebound is a rebound toward mediocrity, not toward contention.

Both appear on an "under-achiever" list. Only the first kind is a team whose record is lying about its quality.

The same asymmetry runs the other way. The Broncos and Jaguars over-performed, but from a genuinely strong efficiency base (+0.179 and +0.194) — they were good teams that were also fortunate. The 49ers at −0.026 and the Panthers at −0.200 were not.

What this says about 2026

Applying the pattern rather than a prediction, since the effect is a tilt:

• Kansas City and Indianapolis are the two teams in the league whose 2025 records most understate how they played. Both combine a positive net EPA with a losing or .500 finish — the exact profile that rebounded hardest in the historical sample.
• San Francisco, Carolina and Pittsburgh carried records their efficiency did not support. Historically, this group doesn't collapse; it converges. The 10-win-plus version of it landed at .573 the following year regardless.
• Denver and Jacksonville are the ones to be careful about dismissing. They over-performed, but their underlying play was strong. Regression from 14-3 to 11-6 is not the same story as regression from 12-5 on negative efficiency.

The honest caveats

• The effect is small. A partial correlation of +0.132 sits at the bottom of the ladder above: once you already know a team's record, its efficiency sharpens the picture of next season only slightly. It shifts the odds a little over many seasons; it does not tell you what one team will do next year.
• One season of EPA is still one season. Seventeen games of per-play data is more stable than seventeen games of win-loss, but it is not stable. The teams above could each be a coaching change, an injury, or a schedule swing away from any of this mattering.
• Rosters aren't constant. This analysis treats a franchise as continuous year to year. Quarterback changes in particular can make a prior-season efficiency number describe a team that no longer exists.
• These are regular-season figures only, 2002–2025, ties counted as half a win, and 2026 rows are excluded from every calculation because the season has no completed games yet.

See it yourself

The full four-way EPA splits for every team and season are on the SpreadTrends Rankings page under the EPA Ratings tab, alongside the records they're supposed to explain. The gap between the two columns is the whole article.

Data: 768 team-seasons 2002–2025 from the SpreadTrends Rankings and Games tables; 736 year-over-year pairs. Expected wins from a same-season fit of win percentage on net EPA (r = .845). Regular season only.