Elo ratings in basketball: the idea and its uses
Elo began as a chess rating and now powers many basketball forecasts. This article explains the formula with a worked example, the adjustments basketball needs, and what an Elo rating can and cannot tell you.
A rating system has to answer one question: given everything we have seen so far, how strong is this team? The most durable answer to that question in competitive games was devised for chess. Arpad Elo, a Hungarian-born American physics professor and chess master, designed a method for rating players that the United States Chess Federation adopted around 1960 and the international federation, FIDE, adopted in 1970. The method is so simple that you can run it in a spreadsheet, and so adaptable that it has been applied to soccer, football, baseball and basketball, and to video games and online matchmaking.
In basketball, Elo ratings have become a standard tool for fans and analysts who want a single, self-updating number for each team that rises with wins and falls with losses, and that can be turned into a win probability for any matchup. This article explains how the system works with a worked example, what has to be added to make it suit basketball, how the ratings can be used, and where they fall short. It sits alongside Net rating: the best single number for a team and Pythagorean wins: using points to predict records as one of the main ways to put a number on team strength, and it feeds directly into the forecasting method in Simulating a season: how win probabilities are built.
The core idea: ratings that learn from surprises
Imagine you know nothing about two teams except that one has a rating of 1600 and the other 1500. The ratings are not points or wins; they are positions on a scale where only differences matter. The 100-point gap means something specific: the higher-rated team is expected to win a certain share of the time. The system's first job is to convert the gap into an expectation.
Elo uses a smooth curve for that. With the usual scaling, a gap of 400 points means the stronger side is expected to win about 91 percent of the time, a gap of 100 points means about 64 percent, and a gap of zero means 50 percent. The formula behind it is a logistic curve: the expected result for team A against team B is one divided by one plus ten raised to the power of the rating gap (B minus A) divided by 400. You do not need to memorize it. You only need to see what it implies, which is that the expectation changes quickly near an even matchup and levels off as the gap grows.
The second job is to learn. After a game, the system compares what happened with what it expected. If the favorite wins as expected, its rating rises a little and the underdog's falls a little. If the underdog wins, the changes are larger, because the result was a surprise. The size of every adjustment is set by a single number called the K-factor. In formula terms, the new rating equals the old rating plus K times the difference between the actual result (1 for a win, 0 for a loss) and the expected result.
A worked example
Take the 1600 and 1500 teams on a neutral floor, and suppose K is 20. The expected probability that the 1600 team wins is about 0.64. If it wins, the actual result is 1, and the gap between the result and the expectation is 0.36. Multiply by K and the winner gains about 7 points, finishing near 1607, and the loser drops by the same amount, to about 1493.
Now suppose the underdog wins. The surprise is larger: the favorite had a 0.64 expectation and got 0, so its rating falls by 0.64 times 20, about 13 points, to roughly 1587, and the underdog rises by 13, to about 1513. The same rule produced two different updates, and they came out in proportion to how unlikely the result was.
Two properties are worth noticing. First, the system is zero-sum: every point the winner gains is a point the loser loses, so the average rating across the league stays fixed. Second, it self-corrects. A team whose rating is too low will keep winning more than the system expects, and the surprises will push its rating up until the expectation matches reality. A team that is too high will do the opposite. There is no need to adjust for strength of schedule separately, because the schedule is already inside the update: beating a strong team earns more than beating a weak one, and losing to a weak team costs more than losing to a strong one. That is the same logic as the schedule adjustment described in Strength of schedule and how to adjust for it, but done game by game.
What basketball needs to add
Chess has no home court, and a draw is common. Basketball has neither, and it has a lot of games and a lot of information in the score. The basic system gets three modifications.
Home-court advantage
Home teams win more than half their games, as detailed in Home-court advantage in the numbers. To account for it, the expectation is calculated with a bonus added to the home team's rating. In public basketball versions the bonus has typically been on the order of 100 rating points, which corresponds to the home side being a roughly 64 percent favorite when the teams are otherwise equal. The bonus does not alter anyone's stored rating; it is added only when predicting and updating that particular game. A road team that beats an equally rated home team gains more than a home team that beats the visitor, because the road win was less expected.
Margin of victory
A win by 30 says more than a win by 1, and a system that treats them alike throws away information. Basketball versions of Elo therefore scale the update by the margin. The multiplier grows with the margin but with diminishing returns, so that the difference between winning by 5 and winning by 10 is larger than between winning by 25 and winning by 30. A good implementation also corrects for a subtle problem: strong teams tend to win by large margins against weak ones, so without a correction they would collect inflated gains for results the system already expected. The standard remedy reduces the multiplier when the winner was already the heavy favorite. Without these adjustments, the ratings chase blowouts. With them, the ratings agree reasonably well with the ones that come out of point-differential systems, which is why the topic is closely tied to Margin of victory and what it says about a team's true level.
Carrying ratings across seasons
Teams change between seasons. Players leave, rookies and free agents arrive, coaches come and go. A team that finished at the top is unlikely to repeat exactly, a point developed in Regression to the mean: why extreme seasons rarely repeat. So at the start of each season, the usual practice is to pull each team's rating partway back toward the league average. A typical choice keeps most of the previous rating, such as three-quarters of the distance from the average, and discards the rest. That keeps good teams good on paper without pretending that a year of change has no effect.
Playoff games are sometimes handled with a different K-factor, since the matchups are repeated and the stakes differ. These choices are tuned by testing which settings produce the best predictions on past games. They are not laws of nature, and the parameter values different authors use vary.
Turning a rating into a forecast
Once a rating is in hand, it supplies a probability for any game. Take the two teams' ratings, add the home bonus to the home side, apply the logistic formula and read off the probability. In a typical basketball setup, a certain number of rating points corresponds to one point of expected scoring margin; versions in common use sit in the neighborhood of 25 to 30 rating points per point of margin. That conversion lets a rating gap be expressed as a point spread, which is useful for comparing the model's view to other sources, or for reading a ratings table the way you would read a spread.
From game probabilities, you can build season forecasts. Feed the whole remaining schedule through the probabilities, simulate the results thousands of times and tally how often each team reaches each win total, seed or round. That is the method developed in Simulating a season: how win probabilities are built. The Elo rating is the engine; the simulation is the vehicle.
How to tell whether a rating is any good
A rating system makes claims that can be checked, which is one of its strengths. The test is calibration. Collect every game in which the model gave one side roughly a 70 percent chance and see how often that side actually won. If the answer is close to 70 percent, the model is well calibrated. If favorites at 70 percent win 80 percent of the time, the model is too timid; if they win 60 percent, it is overconfident. Do the same at every probability level and you have a picture of whether the numbers can be trusted.
There are summary scores for this, such as the Brier score, which averages the squared gap between the stated probability and the outcome, but the intuition is enough. A forecast of 90 percent that fails is penalized heavily, a forecast of 55 percent that fails barely at all, and a model that hedges every game at 50 percent earns a mediocre score because it never says anything. The goal is to be bold when the evidence is strong and cautious when it is not.
This is also how the parameters get chosen. An analyst runs the system over many past seasons with different K-factors, home bonuses and season-to-season carryover, and keeps the settings that forecast the next game best. The caution is that tuning on past games can overfit. A setting that looks perfect on one era may do worse on another, which is a reason to prefer simple, round numbers over finely tuned ones.
Elo and other ways of rating teams
Elo is one of several tools, and it helps to know where it sits among them.
| Approach | What it uses | Main strength | Main weakness |
|---|---|---|---|
| Win-loss record | Results only | Easy to read | Ignores schedule and margin |
| Net rating or point differential | Scores, per possession | Stable, predictive | Needs adjusting for opponents |
| Elo | Results and margins, updated game by game | Self-correcting, simple | Slow to see roster changes |
| Player-based models | Individual ratings and minutes | Sees injuries and trades | Needs detailed data and assumptions |
Each answers a slightly different question. The record tells you what happened. Point differential tells you how well a team played. Elo blends the two with a memory, and a player-based model asks what the team is now, given who is available. The approaches complement one another, and the habit of comparing them is useful: when a team's Elo rating is high but its record is mediocre, it has probably been unlucky in close games; when its record is better than its rating, it has probably been fortunate. The same kind of comparison is made between wins and expected wins in the Pythagorean approach.
The idea travels across leagues. The same mechanics can be applied to the WNBA, to college basketball with its more than 350 teams, and to national teams, where games are rare and a modest K-factor matters because each result carries a lot of weight. College is the hardest of these, because teams from different conferences meet so seldom that the ratings of whole conferences float on a small number of cross-conference games. That is why college systems built on possession-based efficiency, described in College efficiency metrics: KenPom, NET and the ratings used to select the field, carry so much weight there.
Other uses
Elo is also valuable for looking backward. Because it needs only scores and schedules, it can be calculated for every team in a league's history. That makes it a natural tool for the "greatest team ever" arguments that are a basketball tradition. Rank all teams by their peak rating and you get a list, and you can study how long dynasties lasted, how fast ratings fall after a title and how the spread of team strength has changed. The great runs described in Great dynasties: Celtics, Lakers, Bulls, Spurs and Warriors compared show up on such a chart as long stretches at the top.
A caution applies to cross-era comparisons. The ratings are relative: they measure a team against the league of its own time. A 1960s champion and a 2010s champion can have similar peak ratings without saying anything about who would win a game between them, because each rating is measured against a different population of opponents, under different rules and with different athletes. Elo can tell you how far a team stood above its contemporaries. It cannot tell you how the eras compare.
Another use is as a baseline. Because Elo is simple, it provides a benchmark for fancier models. If a complicated model cannot beat the Elo forecast, it is not adding much. Where more sophisticated approaches do better, the gain often comes from information Elo does not use, such as which players are available.
Where Elo falls short
The weakness that matters most in basketball is that Elo knows teams, not players. A rating is a number attached to a franchise, and it updates only after games are played. If a team trades for a star, or its best player is hurt, the rating does not move until the results arrive. A team with its star out for the playoffs may be far weaker than its rating, and a team that has just made a big acquisition may be stronger. In a sport where five players are on the floor, and where a single player can move a team's strength by a large amount, that lag is a real limitation. Later, more elaborate versions incorporate player information to address it, which is a sign of where the field has gone.
A second weakness is that the rating cannot distinguish a lucky stretch from a real change. A team that wins several close games in a row will be pushed up, though a run of narrow results is often noise. The margin-of-victory adjustment helps, but only partly. The cautions in Small samples in basketball: when a hot streak is noise apply as strongly to a rating as to a shooting percentage.
A third is the tuning. The K-factor sets the speed of learning, and no single value is right. A high K reacts quickly to new information but jumps around on noise; a low K is stable but slow to notice a real change. Basketball's long schedule favors a modest K, because 82 games provide a lot of data, but any choice is a trade-off. A fourth is that the formula assumes a single dimension of strength. Real matchups have style effects, and a team's strength against one opponent can differ from its strength against another, which Elo cannot express.
Finally, an Elo rating is a statement about probability, not certainty. A team with a 70 percent chance to win loses three times in ten, and a rating that says so is working properly when that happens.
What to take away
Elo gives each team a rating, turns rating gaps into win probabilities and updates the ratings by the size of each surprise. In basketball it needs a home-court bonus, a margin-of-victory adjustment and a partial reset between seasons, and its value lies in being simple, self-correcting and usable for both forecasting and history. Its blind spots are that it tracks teams, not rosters, and that its tuning involves judgment. Used with those limits in mind, it is a reliable first guess at how strong a team is. You can compare ratings-style views with the current table on the standings page and the ranked lists on the rankings page.