Two NBA teams can have the same record and be in very different shape. One has won eight of its last ten, the other has lost six of its last seven. The standings treat them as equal. Our ELO ratings do not, and that is why our models use them.
What is an ELO rating?
ELO was invented by the physicist Arpad Elo to rank chess players. The idea is simple. When two opponents meet, the winner takes rating points from the loser. How many points change hands depends on how surprising the result was: beating a stronger team earns more than beating a weaker one, and losing to a weaker team costs more than losing to a stronger one.
The 1500 baseline
Every team starts the season at 1500, the league average. Winning teams climb above it, losing teams drop below it. A team at 1600 is playing well above average, a team at 1400 is struggling. Ratings reset to 1500 at the start of every season, so compare teams within one season, not across seasons. The "ELO±" column on our ELO standings page shows how far each team has moved from 1500.
Why the standings are not enough
Take two teams at 25-20 halfway through the season:
Same record, different teams
Team A (25-20):
- Lost 6 of their last 7 games
- Lost those games by 8 points on average
- Beat mostly weak opponents early in the season
Team B (25-20):
- Won 8 of their last 10 games
- Won them by 12 points on average
- Several of those wins came against top teams
The standings say these teams are equal. Their ELO ratings will not be. Team B has been taking points from strong opponents recently, and Team A has been giving them away.
What our ELO adds to the basic version
Classic ELO only looks at who won. Ours takes a few more things from each game into account:
Winning margin
A 25-point win says more about the gap between two teams than a 2-point win. The bigger the margin, the more points change hands.
Home court
Teams play better at home. When we work out the expected result, the home team gets a small head start. A home win against an equal opponent therefore moves the rating less than a road win does, because it was more likely to happen.
Rest days
A team on the second night of a back-to-back is at a disadvantage. The expected result gives a small bonus for each day of rest, capped so that a long break does not count as a big edge.
How much a game can move the rating
This is set by what is usually called the K-factor. Ours is not fixed: it is higher for games between closely rated teams and for big winning margins, and lower when a much stronger team beats a weaker one, which tells us little that we did not already know.
Opponent's win percentage
One of our versions also weighs the opponent's season win percentage, so beating a team that wins 70% of its games counts for more than beating one that wins 30%.
ELO vs. the standings
What each one tells you
League standings
- Total wins and losses
- Conference and division record
- Current playoff position
- Games behind the leader
Our ELO ratings
- How strong a team is now, not just its record
- Who the wins and losses came against
- How big the wins and losses were
- Home and away context
- Rest between games
Nine versions of ELO
We do not compute just one ELO. We compute nine, each switching on a different mix of the adjustments above. They often agree. When they do not, the difference itself is useful information for the models.
| ELO version | What it uses |
|---|---|
| Point Differential | The result and the winning margin |
| Point Diff + Dynamic K | Same, with the rating change scaled by the margin and by how close the two teams were |
| Point Diff + Damped | Same, with every rating change scaled down, so ratings move more slowly |
| Point Diff + Home Adjusted | Adds home court to the expected result |
| Point Diff + Rest Adjusted | Adds rest days to the expected result |
| Point Diff + Fully Adjusted | Home court and rest days together |
| Traditional ELO | Win or loss only, the classic chess formula, with the dynamic K-factor |
| Traditional + Fully Adjusted | Our main rating, used for the power rankings: classic ELO with home court, rest days and the dynamic K-factor |
| Win Percentage Weighted | Weighs each game by the opponent's season win percentage |
Why the main rating is Traditional + Fully Adjusted
In our tests it gave the most stable rankings: it rewards wins, adjusts for home court and rest, and reacts more to close matchups and big margins. It is the rating you see on our ELO Power Rankings page.
How ELO feeds our predictions
ELO does two jobs. It drives the ELO Power Rankings, and it is one of the inputs to our prediction models:
- After each game, all nine versions are updated for both teams.
- The power rankings update from the main version, with conference standings, streaks and progression charts.
- Before each game, the ratings of the two teams go into the models alongside player statistics, the expected starting lineups and recent form.
- The models turn all of this into win probabilities, point totals and spreads.
The models are trained on every NBA game since the 1996-97 season.
Note: ELO is one input among many. Basketball stays unpredictable, and no rating removes that.
See the ratings
Browse the live ELO power rankings for all 30 NBA teams, or see the predictions they feed: win probabilities, total points and spreads.
In short
A team's record tells you about its season. Its ELO rating tells you how strong it is now: who it has been beating, by how much, and in what conditions. The progression chart on the power rankings page shows how each team's rating has moved game by game, so teams on the rise or in decline stand out.