Team comps: what 12 lookalike teams can and can't tell you

"They look like the 2005 Colts" is one of the oldest moves in football talk. Every team reminds someone of a team from the past, and the comparison does a lot of work: it tells you how good a team is, what kind of good, and how the story might end.

We built a page that does this systematically. Pick any NFL team and any point in its season on the Team comps page, and it finds the 12 team-seasons since 1999 that played most like it after the same number of games. Then it shows how each of those seasons finished.

We also tested it, and you should know the result before you use it: the comps do not predict how a season will end any better than a simple formula using the same numbers. They also don't beat the team's current record. Here is what we built, what the test found, and what the comps are still useful for.

How a team gets matched

Every team gets three numbers after every game:

• Offense and defense: expected points added per play on passes and runs, adjusted for the opponents faced and for home field. Snaps from blowouts are left out (win probability under 5% or over 95%).
• Special teams: expected points added per game on kickoffs, punts, field goals and extra points. This one is not adjusted for opponents.

Each number is expressed against every team at the same point since 1999, so +1.0 means about one standard deviation better than average (roughly the league's top sixth). Teams are lined up by games played, not by calendar week. A team three games in is compared only with other teams three games in, which keeps byes from mixing a two-game sample with a three-game one.

The 12 comps are the earlier team-seasons closest on those three numbers. Special teams counts half, because it is the least stable of the three from year to year. We set these choices once, before testing, and never tuned them. Final records from 16-game seasons are scaled to 17 games so all eras sit on one scale.

What the test found

Before running anything, we wrote down the test: for every team in the 2006 to 2025 seasons, after 2 through 10 games played, use the comps to predict the rest of the season, using only seasons before the one being predicted. That's 640 team-seasons and 5,760 predictions. We compared the comps against two simple baselines and ran the test once.

The miss is the average gap between predicted and actual winning percentage over the rest of the season. With 14 games left, 16.7 points is about 2.3 wins. The formula beat the comps at every point from 2 games to 10, and the gap held up when we resampled the seasons (95% interval: 0.4 to 0.9 points). On the playoffs, the comps did clearly worse than simply knowing the record. For scale, giving every team the league's base playoff rate scores about 0.239 on the same measure, and lower is better.

The reason is not mysterious. Twelve neighbors in a three-number space are a noisy version of a fit over every past team. And the comps ignore the record on purpose. A 3-0 team and a 1-2 team with the same ratings get the same comps, even though the 3-0 team already has the wins banked. That costs the most on playoff odds, where banked wins matter a lot.

What comps are good for

Seeing the range, not just the middle. An average of 12 teams is a weak forecast, but the spread of 12 teams is honest about how wide the outcomes run. Across 2006 to 2025, three games in, the best and worst of a team's 12 comps finished a median of 8.5 wins apart. Teams that looked alike after three games ended up anywhere from bad to very good.

Miami this year is the clearest example. The Dolphins are 0-3, with the worst defense in the league by this measure (−1.53). Eleven of their 12 comps finished between 2-14 and 7-9. The twelfth is the 2025 Patriots, who went 14-3 and reached the Super Bowl. The average says 5.8 wins. The full list says it is very likely bad, and a turnaround is not impossible.

Seeing what kind of team it is. The match is on offense, defense and special teams separately, so comps share a style as well as a level. San Francisco's offense is at +2.66, far beyond anything else this year. Its comps are a list of great offenses: the 2001 Rams, 2009 Colts, 2010 and 2011 Patriots, 2019 and 2020 Chiefs. Eleven of the 12 made the playoffs and six reached the Super Bowl. Minnesota, also 3-0, is the opposite shape (offense −0.97, defense +1.57), and its comps are defense-first teams like the 2015 Broncos and 2008 Titans. Half of them made the playoffs. The page also shows the passing and running splits for context, though they don't pick the comps.

Checking a narrative. When a team starts 3-0 and the talk turns to contender, the comps show what history did with teams that actually played like that, which is often different from teams that merely won like that.

What comps shouldn't be used for

A win total or a playoff probability. This is the main one. The comps' average is the worst of the three methods we tested, and on playoffs they lose to the record alone. For a forecast, use Super Bowl odds, which come from the season simulation, or our win totals.

Betting. The comps lose to a formula built from public numbers. A betting market already prices those numbers in, so there's no reason to expect the comps to beat it.

One comp as the answer. "They look like the 2005 Colts, who went 14-2" picks the best line out of twelve. Chicago and Las Vegas share 10 of their 12 comps this year, and that list includes both the 14-2 Colts and the 4-12 Cowboys of 2015. Quoting either one alone tells you more about the person quoting than about the team.

Ranking teams. Comps answer "who did this look like," not "who is better." Team ratings do the ranking.

Anything the three numbers don't see. The match knows how a team has played so far. It doesn't know about a quarterback injury, a trade, or the schedule ahead. Special teams isn't opponent-adjusted, and after two or three games every number rests on very few plays.

Where this leaves it

We built this expecting the comps might add something, tested it the way we said we would, and the answer was no. We kept the page because the comparisons are still interesting and the range of outcomes is useful. The test result sits at the top of the page so nobody mistakes it for a forecast.

Two things were left untested, and each would need its own test written down before running: adding the current record to the match, and tuning the number of comps or the weights. With the formula already ahead using the same inputs, we don't expect either to close the gap.

Pick any team and any point in its season on the Team comps page. The figures above are 2026 through Week 3.