NBA · 2026-27 · Week 2
The Early-Season Mirage
Four games into an NBA season, a team's is mostly noise. We replayed every game of the last five seasons to measure when the numbers start telling the truth.
Every team starts as a rumor
Net rating is the point differential per 100 : the best single number for how good a team is. After one game last season, it ranged from −33 to +32. By April, last season's 30 teams fit between −11.5 and +10.9.
The shaded funnel shows how far an exactly average team would drift by luck alone. Early on, almost every team is inside it.
Net rating after each game, 2025-26 (all 30 teams)
Pick a team to follow its path. Hover to read any team.
Net rating = 100 × (points − opponent points) ÷ possessions, from cumulative totals. Source: NBA Stats API.
When the numbers start to mean something
To test how much an early rating really says, compare a team's rating after its first k games with how it played in all the games that followed. Comparing against the full-season number would cheat, because the early games are part of the final total.
Averaged over five seasons, net rating reaches 0.5 by game 6 and 0.7 by game 14. Seasons differ: 2024-25 got there by game 7, while 2021-22 never did. The curve then levels off around 0.7, because teams genuinely change during a season through injuries, trades and late-season tanking.
Defense is the noisiest of all. barely gets past 0.5 on average, so early "elite defense" headlines deserve the most skepticism.
How well the first k games predict the rest of the season
Correlation across 30 teams. Five-season average in bold, each season faint. Hover for values.
Correlation of each team's rating over games 1–k with its rating over games k+1–82, regular seasons 2021-22 to 2025-26. Source: NBA Stats API.
Early leaders, late fades
Rankings tell the same story. Chicago was 5th after four games and finished 22nd. was 20th and finished 3rd. Some early signals were real: San Antonio was 1st after four games and finished 2nd.
Net-rating rank after 4 games vs. final rank, 2025-26
Teams that moved 10 or more places are colored. Hover a line for details.
Last season knows more than you think
Early games aren't the only evidence. Last season's net rating, before any offseason moves are accounted for, is a steady predictor all year. For the first five games, it predicts the rest of the season better than the new season's own results do. The current season's games overtake it around game 6.
Using both beats either one alone, and that's how the estimates below are built.
Predicting rest-of-season net rating: this season, last season, or both
Correlation with games k+1–82, pooled across 2022-23 to 2025-26 (120 team-seasons). Hover for values.
So what do we know about this season?
Four games in, the of each team's rest-of-season net rating leans more on last season than on this one. Chicago's −17 start becomes about −5.5, close to where it finished last year. Minnesota's +16 start nudges its estimate from last season's +3.2 to about +4.5.
The ranges are wide on purpose. They describe how the rest of the season could realistically go, and they don't know about trades, injuries or new coaches.
2026-27 net rating so far, and the best estimate for the rest of the season
Regular season through Oct 27 (56 games). Hover a team for details.
Show as a table
| Team | Games | Net so far | 2025-26 | Best estimate | 90% range |
|---|
Check back around game 14. Over the last five seasons, that's when a team's own results became a good guide to the rest of its year.
Method
- Data: every regular-season game from 2021-22 to 2025-26 (6,150 games) and 2026-27 through Oct 27 (56 games), from the NBA Stats API via the open-source
nba_apiclient. 2019-20 and 2020-21 are excluded (bubble, shortened schedules, empty arenas). - Ratings are built from cumulative totals, never averages of single-game ratings. Possessions are estimated as FGA + 0.44 × FTA − offensive rebounds + turnovers. Against the NBA's own possession counts for 2025-26, season net ratings differ by at most 0.15 and every team's rank is identical.
- Predictive test: Pearson correlation across the 30 teams between games 1–k and games k+1–82, so the early games never appear on both sides.
- Noise model: a team's single-game net rating varies by about 14.4 points (standard deviation), while true team quality varies by about 5.0. Signal and noise are equal at about 8 games, consistent with the correlation curve.
- Best estimate: a linear regression of rest-of-season net rating on (a) net rating so far and (b) last season's final net rating, fitted at the same game count across four season pairs (120 team-seasons). The 90% range comes from the regression's typical error.
- Limits: five seasons of history. The estimate doesn't know about offseason trades, injuries or coaching changes, and makes no adjustment for opponents or rest. The highlighted paths and rankings use 2025-26 as the example season.