← Learn

12 min read · updated September 26, 2026

What is Elo?

The rating system behind half the numbers on this site, from a chess federation in the 1960s to your Sunday picks: what a rating is, how it turns into a win probability, how it moves after every game, and what it gets wrong.

The short version

Every NFL team has one number, its Elo rating. The average team sits at 1505. Right now the best team in the league is about 1655 and the worst about 1354. When two teams play, the gap between their numbers becomes a win probability: the bigger the gap, the surer the favorite. After the game, the winner takes some points from the loser, more if the result was a surprise and more if it was a blowout. That's the whole idea. Everything else is detail, and the details are what this article is about.

Where it comes from

Arpad Elo was a physics professor and a chess master. In the 1960s he designed a rating system for the United States Chess Federation to replace one that nobody trusted, and it spread to almost every competitive game: world chess, Go, esports, online matchmaking. The idea that made it last is simple. A rating is a prediction. If your rating says you should beat someone three times in four, and you do, your rating should barely move. If you lose, it should drop, because the prediction was wrong.

FiveThirtyEight adapted Elo to the NFL in 2014, added a few football-specific pieces (home field, margin of victory, and later the starting quarterback), and published a forecast for every game through the 2022 season. Delta Sunday picks up from their final published ratings and keeps the model running: the same core method, a couple of constants retuned, and a quarterback adjustment that updates every three hours from the depth chart.

Every team is a number

A rating means nothing on its own; only the gap between two ratings matters. A 1600 team isn't "good" because 1600 is a big number. It's good because it's about 95 points above the average team, and 95 points turns out to mean something specific about how often it wins.

Ratings are zero-sum: every point one team gains in a game, its opponent loses. So the league average stays put at about 1505 no matter what happens. The only way for a team to climb is to take points from the teams it plays.

From ratings to a win probability

This is the one formula worth knowing, because every Elo number on the site comes from it:

home win probability = 1 / (1 + 10 ^ (−gap / 400))

The gap is the home team's rating minus the away team's, plus adjustments. The 400 sets the scale: a gap of 400 points means the stronger team is ten times as likely to win as to lose, about 91%. A gap of zero is a coin flip.

The adjustments that go into the gap:

• Home field: +48. Home teams win more often than their ratings alone say. On a neutral field, like a game in London or the Super Bowl, this is zero.
• Coming off a bye: +25 for the rested team.
• The quarterback: covered below.
• Playoffs: the whole gap is multiplied by 1.2, because better teams win a little more often in January than the regular-season numbers suggest.

Two examples. Two average teams, one at home: the gap is 48, so the home team is 57%. A 1600 team at home against a 1500 team: the gap is 100 + 48 = 148, so the home team is 70%.

From a probability to a point spread

Betting lines are quoted as point spreads, so it helps to translate. A useful rule: 25 Elo points is about one point on the scoreboard. The 148-point gap in the example above is roughly a 6-point favorite. You can see this on every game page, where Elo's implied spread sits next to the sportsbooks' line.

After the game: how ratings move

The update rule is where Elo learns:

change = K × margin multiplier × (result − expected)

• Result is 1 for a win and 0 for a loss (a tie is 0.5).
• Expected is the win probability from before the game.
• K is 20. It sets how fast ratings move. Too high and a team swings wildly on one game; too low and the ratings take half a season to notice a team got better.
• The margin multiplier grows with the margin of victory, but slowly (it uses a logarithm), and it grows less when a big favorite wins big. Beating a bad team by 30 should count for more than beating it by 3, but not ten times more, and a heavy favorite running up the score is less informative than an underdog doing it.

Three results for the same game, the 1600 team at home against the 1500 team, with the home team at 70%:

The upset moves ratings almost three times as much as the expected win. That asymmetry is the whole point: the system barely learns from what it already believed, and learns a lot from what it didn't.

The quarterback

The biggest thing a plain team rating misses is who is playing quarterback. A team rating built on twelve weeks of one starter is wrong the week a backup takes over. So the model keeps a running value for every quarterback, built from what he does in each game: completions, passing and rushing yards, and touchdowns count for him; pass attempts, runs, interceptions and sacks count against him, so efficiency matters more than volume. It keeps the same running value for each team's quarterback play overall.

The quarterback adjustment is the difference between the two, converted to Elo points. If a team's usual starter plays, the adjustment is near zero, because he is the team's usual quarterback play. If a backup starts, it goes negative. If a clearly better quarterback takes over, it goes positive. On the site you'll see it next to the name, like Murray (+100 Elo): that's four points of spread, already inside the Elo number. The adjustment is capped at 100 either way.

Rookies start from a value based on where they were drafted, since there's nothing else to go on. Every three hours, the site reads each team's depth chart and injury report to work out who is expected to start, so a benching or an injury shows up in the forecast before kickoff, not after the box score.

We checked whether adjustments this large are too much by replaying the model over every season since 1999. At full strength the adjustment is worth about 51 points a season over ignoring quarterbacks entirely. When it's big, 70 points or more (744 games), the side it favored won 58% of the time; the model said 57%, the market 53%. Big quarterback swings are real, and the model sizes them about right.

Between seasons

Rosters change every spring. Stars retire, draft picks arrive, coaches get fired. So before each season, every rating is pulled halfway back to 1505. A 1650 team starts the next season at about 1578; a 1400 team starts at about 1453. FiveThirtyEight pulled one third of the way. We tested both on recent seasons, and halfway did better, which fits a league where teams rise and fall faster than they used to.

How good is it?

Honestly: useful, but not as good as the betting market. Scored the way this game scores your picks, over 6,918 games from 1999 to 2025, the closing betting line averages about 998 points a season. FiveThirtyEight's Elo, the version ours descends from, averaged 874 over 1999 to 2022. Our own version, replayed from 1999 with exactly the code that runs today, averages 839 over all 27 seasons, about 55 a season behind theirs over the seasons both cover. Their model adjusted each quarterback for the defenses he faced and pulled quarterback ratings back between seasons; ours doesn't yet. Those are the likely reasons for the gap, and the next thing we'll test. This article will change when the numbers do.

The market is better because it knows everything Elo knows and a great deal more: injuries below the quarterback, weather, motivation, and the opinions of people who bet serious money on being right. Elo knows scores, home field, rest and the quarterback.

Where Elo is most wrong is exactly where it disagrees most with the market. When the two are ten or more points apart, Elo's side has won about half the time over 1,102 games, while Elo said 65%. When Elo and the market disagree, the market has usually been right. (That's what the bold call of the week is: the game where Elo disagrees most with the market. It hits less than half the time.)

So why keep Elo at all?

• It explains itself. Every number traces back to results and a quarterback. A market price just is.
• It exists before the market does. Ratings forecast every game on the schedule months ahead, which is what the playoff odds and the survivor planner run on.
• It's independent. It doesn't know what the market thinks, so where they agree, that's two different kinds of evidence pointing the same way. When they agree tightly on a favorite, favorites have won a little more often than either one said.
• It's the starting point for The System. The System starts from the market and adds Elo at a quarter weight, plus a few rules that held up over 27 seasons. It beats Elo clearly and runs about even with the market.

Where to see it on the site

• Ratings: every team's current number and how it got there.
• QBs: every expected starter and his adjustment.
• Any game page: Elo's win probability, its implied spread, and both quarterbacks.
• Predictions: Elo next to the markets for every game, and which games they disagree on.
• How it works and The System: how Elo fits into the forecast you're trying to beat.

A few terms

• Rating: a team's strength in Elo points. The average is 1505.
• Gap: the difference between two ratings, after home field, rest and quarterback.
• K: how fast ratings move after a game. Here, 20.
• Reversion: pulling every rating halfway back to average between seasons.
• Implied spread: Elo's gap turned into points, about 25 Elo points to one point.