How it works
Every team gets one number, AdjEM, that says how much better or worse it is than an average NBL team. The power rankings and every prediction on the site come from it.
What the rating means
AdjEM is how many more points a team scores than it gives up every 100 possessions, against an average team on a neutral court. 0 is average. A +5 team would beat an average team by about 4.0 points in a normal game of around 81 possessions each.
A possession is one trip down the court. It ends with a score, a missed shot the other team rebounds or a turnover. Counting per possession lets a fast team and a slow team be compared on the same scale.
| AdjO | Offence: points scored per 100 possessions. Higher is better. |
| AdjD | Defence: points allowed per 100 possessions. Lower is better. |
| AdjEM | AdjO minus AdjD. Higher is better. |
| AdjT | Tempo: possessions per 40 minutes. Higher is faster. |
"Adj" means adjusted for the opponents each team has played. All the terms are in the glossary.
How the ratings are worked out
- Count possessions. From each box score: field-goal attempts, minus offensive rebounds, plus turnovers, plus 0.44 × free-throw attempts.
- Allow for the opponent. Scoring 110 against the best defence in the league is worth more than 110 against the worst. The model looks at every game at once and finds the ratings that best explain all the results, given who played whom and where. Home court is worked out at the same time. It's the same method Ken Pomeroy uses for US college ratings (a ridge regression).
- Count recent games more. A game's weight halves every 120 days, so a game from four months ago counts about half as much as last night's.
- Start from last season. A handful of games is mostly luck, so last season's games are used early on. They fade out as this season goes and are gone once 60% of the regular season is played (around Christmas).
- Hold back early. Every rating is pulled a little towards average until there's enough evidence, so a team that wins its first three isn't rated a champion.
I tried other weightings on past seasons and none predicted much better.
How games are predicted
Score. Each team's offence is set against the other team's defence, at the pace both usually play, plus home court. That gives an expected score for each side.
Win chance. Real results land within about 13.8 points of the predicted margin two games in three. A team favoured by more than that wins most of the time. A 1-point favourite is close to a coin flip.
Season odds. The rest of the season and the finals are played out 10,000 times. In each run every team's rating gets a small random nudge up or down, because the ratings are only estimates. The size of the nudge came from past seasons and it shrinks as the season goes. Ladder and finals odds are the share of runs where it happened.
Missing players
A regular (3 of the team's last 5 games, 12+ minutes a game) who missed the team's last game is assumed to miss the next one too. The prediction then moves by how much better that player is than the teammates who covered the minutes, using each player's Impact and usual minutes. On 2023-24 onwards this made predictions better when a key player was out.
About an hour before tip-off the NBL posts each team's game-day list. The game page shows a light for each player the model has as out:
| Green | Named in the list. Expected to play, so the adjustment comes off. |
| Red | Not named. Out. |
| Yellow | List not out yet. Questionable. |
The model only knows what box scores show, so a new injury isn't picked up until a game has been missed. About a third of these players end up playing, so the adjustment is set at half the size that fits best and capped at 4 points per team per game. On past seasons that was just as accurate and stopped big swings.
League right now
| Average points per 100 possessions | 116.7 |
| Average possessions per 40 minutes | 80.7 |
| Home-court advantage | +1.8 points per game |
| Typical miss on the margin (two games in three are closer) | 13.8 points |
Backtest
To test the model I replayed every season from 2016-10-06 to 2026-11-15 a week at a time. Each week it only knew about games already played, then predicted the next week. That's 1,518 predictions made without knowing the result.
Log loss scores the win chances. Being confident and wrong costs the most. Lower is better, and a 50-50 guess every game scores 0.693.
Calibration
When the model makes a team a 65% favourite, does it win about 65% of the time?
| Favourite's chance | Games | Model said | Won |
|---|---|---|---|
| 50-60% | 733 | 55.0% | 56.1% |
| 60-70% | 578 | 64.3% | 64.2% |
| 70-80% | 178 | 73.7% | 72.5% |
| 80-90% | 28 | 83.3% | 82.1% |
| 90-100% | 1 | 92.3% | 100.0% |
By season
| Season | Games | Winner % | Log loss |
|---|---|---|---|
| 2016-17 | 115 | 60.9% | 0.703 |
| 2017-18 | 118 | 61.9% | 0.662 |
| 2018-19 | 120 | 57.5% | 0.671 |
| 2019-20 | 133 | 66.9% | 0.622 |
| 2020-21 | 171 | 61.4% | 0.658 |
| 2021-22 | 148 | 63.5% | 0.643 |
| 2022-23 | 153 | 58.8% | 0.660 |
| 2023-24 | 154 | 59.7% | 0.670 |
| 2024-25 | 159 | 64.8% | 0.659 |
| 2025-26 | 179 | 65.4% | 0.612 |
| 2026-27 | 68 | 48.5% | 0.719 |
Prediction ledger
Every prediction is saved before tip-off and marked afterwards, so the record can't be changed later. Tracking since 2026-10-14: 39 games marked, winner picked 51.3%, log loss 0.735, average margin error 11.1 points. 97 predictions waiting on games.
| Game | Predicted | Result |
|---|---|---|
| Tasmania @ Brisbane 2026-11-15 | TAS 55%, by 1.6 | 104–92 |
| NZ Breakers @ Sydney 2026-11-15 | SYD 75%, by 9.3 | 102–111 |
| Illawarra @ Perth 2026-11-14 | PER 61%, by 3.8 | 100–111 |
| SE Melbourne @ Melbourne 2026-11-14 | MEL 64%, by 5.0 | 95–102 |
| Sydney @ Cairns 2026-11-13 | SYD 72%, by 8.3 | 82–84 |
| Brisbane @ NZ Breakers 2026-11-13 | NZB 65%, by 5.3 | 90–94 |
| Tasmania @ Perth 2026-11-12 | TAS 51%, by 0.4 | 92–98 |
| Melbourne @ Adelaide 2026-11-12 | MEL 56%, by 2.2 | 93–101 |
| Illawarra @ SE Melbourne 2026-11-11 | SEM 71%, by 7.7 | 120–110 |
| Cairns @ Perth 2026-11-08 | PER 65%, by 5.1 | 100–92 |
| SE Melbourne @ Sydney 2026-11-08 | SYD 66%, by 5.6 | 99–110 |
| Brisbane @ Melbourne 2026-11-07 | MEL 76%, by 9.9 | 92–120 |
| Illawarra @ Adelaide 2026-11-07 | ADL 60%, by 3.5 | 78–106 |
| NZ Breakers @ Tasmania 2026-11-06 | TAS 60%, by 3.4 | 77–79 |
| Illawarra @ Cairns 2026-11-05 | CNS 55%, by 1.6 | 107–95 |
| Perth @ Brisbane 2026-11-04 | PER 59%, by 3.3 | 104–109 |
| NZ Breakers @ Melbourne 2026-11-02 | MEL 58%, by 2.8 | 74–94 |
| Brisbane @ Adelaide 2026-11-01 | ADL 66%, by 5.7 | 96–89 |
| Tasmania @ SE Melbourne 2026-11-01 | SEM 56%, by 2.2 | 98–89 |
| Illawarra @ Perth 2026-10-31 | PER 70%, by 7.2 | 105–98 |
| Melbourne @ Cairns 2026-10-31 | MEL 61%, by 3.8 | 113–97 |
| Sydney @ Tasmania 2026-10-30 | SYD 58%, by 2.9 | 71–68 |
| Adelaide @ NZ Breakers 2026-10-30 | NZB 63%, by 4.6 | 108–101 |
| Cairns @ Brisbane 2026-10-29 | BRI 58%, by 2.7 | 107–85 |
| Sydney @ Perth 2026-10-28 | SYD 59%, by 3.1 | 99–91 |
| Sydney @ Melbourne 2026-10-25 | SYD 63%, by 4.5 | 95–105 |
| NZ Breakers @ Cairns 2026-10-25 | NZB 65%, by 5.2 | 84–86 |
| Adelaide @ SE Melbourne 2026-10-24 | SEM 56%, by 1.9 | 94–108 |
| Brisbane @ Illawarra 2026-10-24 | ILL 72%, by 8.1 | 98–81 |
| Melbourne @ Perth 2026-10-23 | PER 54%, by 1.5 | 102–103 |
This season so far, replayed the same blind way: 68 games, winner 48.5%, log loss 0.725, margin error 10.9.
Data
Box scores and schedules come from the NBL's official stats feed, mostly through Jason Zivkovic's open-source nblR project. Without it this site wouldn't exist. Games nblR hasn't picked up yet are read from the same NBL feed directly, and the site rebuilds every hour. Fast-break points aren't in this season's feed, so they're not shown.