Our Numbers and Everyone Else’s

Every number on this site starts from one rating, ours. Before you trust any of it, you should know how that rating does against everyone else’s, including on the exam where it had home court and the one where it didn’t.

So here’s the whole thing: what each of the public numbers counts, three exams of which ones predict, and the players where ours disagrees with everybody else’s.

Standing on shoulders

Smart people have been trying to boil a basketball player down to one number for about forty years, and every one of them was right about something. Ours is built on their work, so here it is in order.

Counting the box score. The first tries just added up the stat line. Martin Manley’s Efficiency Rating, from his Basketball Heaven books in the late 1980s, eventually became the NBA’s official EFF. John Hollinger’s PER, in the early 2000s, adjusted box-score production for pace and playing time. Win Shares, from Basketball-Reference, borrowed Bill James’s idea of crediting individual players for team wins and built it around Dean Oliver’s offensive and defensive ratings. The catch with all of them is that they’re ultimately limited to what appears in the box score.

Watching the scoreboard. The other road started with Harvey Pollack, the 76ers’ statistician, tracking plus-minus as early as 1974. The trouble with raw plus-minus is that a bench player who shares the floor with a star looks like a star. Around 2000, Wayne Winston and Jeff Sagarin addressed that with adjusted plus-minus, which tries to separate a player from the teammates and opponents around him. Mark Cuban brought their system to the Mavericks after buying the team, reportedly for more than $100,000 a year.¹ Dan Rosenbaum published the first public version in 2004. In 2010, Joe Sill showed how regularization could tame its instability. That became RAPM.

Doing both. The last decade has mostly been about combining the two: use the box score as a starting guess, and let the scoreboard correct it. Daniel Myers’s BPM, fitted to plus-minus, arrived on Basketball-Reference in 2014. Jeremias Engelmann and Steve Ilardi introduced Real Plus-Minus at ESPN the same year. Then came Jacob Goldstein’s PIPM (2018), FiveThirtyEight’s RAPTOR (2019), Kostya Medvedovsky’s DARKO, a projection model updated every day (by 2019), Taylor Snarr’s EPM (2020) and BBall Index’s LEBRON. Each generation found something the last one was missing.

Seeing the floor. Not every leap was a number. Kirk Goldsberry’s CourtVision, presented at the MIT Sloan conference in 2012, put every shot on a map, and once you could see where the points came from, you couldn’t unsee it. In 2013-14 the league put SportVU tracking cameras in every arena, following every player and the ball through every game, and Second Spectrum took over in 2017-18. That tracking data now feeds the starting guesses in RAPTOR and EPM. Ours uses it for the numbers you see on this site, which look back on a season, but not for the forecast that sits the exams below: there, a simpler starting guess built from the box score alone predicted the next season better.²

Ours belongs to the doing-both family. It isn’t a new idea; it’s a new attempt at an old one, which is exactly why it has to sit the same exams.³

The exams below grade rates, not totals. A rate is how good a player is per possession; a total multiplies that by how much he played. mPM is a rate and mWAR is its total, counted above the kind of player any team can sign for the minimum.⁴ VORP is BPM’s total, and Win Shares is a total too, so they sit the exams as their rates.

That keeps the exams fair, and it leaves out part of what ours does. Outside these exams, our projections, the ones in the Trade Lab, also forecast availability: how much a player is likely to actually play. None of the other rate stats tries to, so every metric gets handed the minutes players really played and is graded on the rate alone. When you want to know who the best player is, though, whether he’ll be on the floor matters a lot. A great player in street clothes is worth nothing that night. If you’re picking an MVP, use mWAR: it counts the games he actually played.

Three exams

Three exams, same rules for everybody. Each team’s prediction is its players’ numbers from the season before, weighted by the minutes they played. That holds for the playoffs too: the 2024 playoffs are predicted from 2022-23 numbers, not 2023-24’s, because several of these metrics fold a spring’s playoff games into that season’s number, and a metric shouldn’t be graded on games it has already seen.⁵⁶

1. Regular season: how close it comes to each team’s scoring margin.
2. Playoffs: how often it picks the series winner.
3. Titles: how much of a chance it gave the champion.

Our number is also entered in pieces, the scoreboard part alone (plain RAPM) and the box-score part alone (our starting guess), with raw plus-minus as the floor. If the whole isn’t better than its parts, you should know.

About that home court. We built our number by running it through the first exam over and over and keeping whatever scored better.⁷ That flatters our regular-season score. We may not be the only ones: anything built to forecast, DARKO included, was presumably tuned on these same seasons too. Either way, we never tuned anything on the playoffs, which makes those the fair fight.

The regular season

Here’s the first exam: 390 team-seasons, 2013-14 to 2025-26.

The typical miss in each team’s schedule-adjusted scoring margin and raw point differential, in points per game, and in its win total (lower is better). *RAPTOR and PREDATOR: 2014-15 to 2022-23 only; over those seasons ours misses by 2.56, 2.67 and 6.9. *RPM: 2014-15 to 2023-24 only; over those seasons ours misses by 2.62, 2.72 and 7.0.

Ours has the smallest miss against every one of them, on all three grades. On the margin, though, the gap to LEBRON is small enough to be luck, ours only just clears that bar against BPM, RAPTOR, DARKO and RPM, and against the rest it isn’t close.⁸ DARKO and LEBRON are the toughest: on the margin ours missed by less in only seven of thirteen seasons against each. Wins are the noisiest grade, since close games swing them, and there none of DARKO, LEBRON and RPM is separable from ours.⁹ Ours doesn’t catch everything, either: whichever grade you use, adding BPM, Win Shares per 48, PER, DARKO, LEBRON and RPM on top of ours still helps a little.¹⁰

Further back

Thirteen seasons isn’t much of a sample, and DARKO is close enough that a couple of odd years could flip it. DARKO publishes every player’s history on darko.app, FiveThirtyEight left RAPTOR’s in its data archive, and Basketball-Reference has BPM for every season, so the first exam can run a lot further back. Ours sets the limit: each prediction pools the five seasons before it, and the play-by-play it reads starts in 1996-97, so 2001-02 is the first season it can predict on a full set. Same exam, same rules, 2001-02 to 2025-26:¹¹

The typical miss in each team’s schedule-adjusted scoring margin, in points per game (lower is better).

Against DARKO, ours has the smaller miss in all four eras, and through 2020-21 the gap is never bigger than luck: ours missed by less in 12 of 20 of those seasons. Since 2021-22, ours missed by less in four of five, a margin luck alone would produce 6% of the time. That’s close to the bar, not over it. Every metric misses by more in those seasons, DARKO most of all, enough to fall behind BPM, and I can’t tell you why. Over all 25 seasons, ours missed by less in 16, which luck alone would produce 1% of the time.

The box-score numbers fall furthest behind in the oldest seasons. From 2001-02 to 2005-06, ours missed by less than RAPTOR and BPM in all five seasons, beyond luck against both. Against RAPTOR it’s ahead beyond luck again from 2013-14 to 2020-21; everywhere else it’s ahead within luck. Win Shares, PER and the two halves of ours trail in every era.

The playoffs

Now the fair fight. We never tuned on the playoffs, and as far as I know nobody else did either, so whatever a metric knows here, it really knows. Every series from 2013-14 to 2025-26, 195 of them, each team weighted by the minutes its players actually logged in the playoffs:

“Chance it gave the winner” is the typical probability each metric put on the team that won the series; a coin flip is 50%. *RAPTOR and PREDATOR: 2014-15 to 2022-23 only; over those seasons ours puts 60% on the winner. *RPM: 2014-15 to 2023-24 only; over those seasons ours puts 59% on the winner.

For a baseline, look at the standings. Last season’s records pick 60% of series right; ours, on the same season-old information, picks 75% and puts 59% on the team that wins, level with the best of the metrics that cover every season, and every metric here clears that bar. This season’s records are a tougher one: they pick 72%, and ours, rerun on the same regular season, picks 71%. Dead even. Back on the main exam, only four metrics are clearly less confident than ours in the winner: the two box-score classics, Win Shares per 48 and PER; raw plus-minus, which doesn’t adjust for teammates; and LEBRON, which is rated one season at a time, so a season-old LEBRON is a stale one. On this season’s numbers, LEBRON draws level.¹² Against everything built on, or fitted to, the adjusted scoreboard, the gap is small enough to be luck.¹³ Among the metrics that cover every season, the best playoff score belongs to our own box-score starting guess, by two series out of 195, which is within luck too.

Why is it so hard to beat the standings? Partly because the playoffs are their own animal. Rotations shrink: on playoff teams, the five players who play the most get 56% of the minutes in the regular season and 68% in the playoffs, more in every season since 2013-14, and the busiest player goes from 66% of the game to 80%. These exams already weight by the minutes actually played. What a rating built over 82 games can miss is how a player holds up when an opponent spends a week game-planning for him. That’s why the Trade Lab doesn’t price a title off wins alone: its title odds add the strength of a roster’s best player on top of its projected wins.¹⁴

Then the titles, which are what everyone wants and what can tell you least: there have only been thirteen in this stretch. Ours had the eventual champion as the strongest team six times, as often as any metric, tied with plain RAPM, raw plus-minus and DARKO. Plain RAPM, BPM, RAPTOR, PREDATOR and RPM gave the champion more of a chance than ours did.¹⁵ For comparison, the team with last season’s best record won three of them, and the team with this season’s best record won five. Thirteen titles can’t separate any of these, so I’m reporting it, not leaning on it.¹⁶

RPM comes with an asterisk of its own. ESPN changed its recipe in 2019-20, adding tracking data, so the RPM in these tables is the original through 2018-19 and the new version after, stitched together. There’s no visible jump where they meet.¹⁷

One metric I wanted isn’t in any of this. EPM’s history sits behind a data feed that Dunks & Threes asks automated readers to leave alone, so we did. Dunks & Threes, who designed the first exam, ran it on EPM themselves: on their version, EPM had the smallest miss of everything they tried, RPM and RAPTOR included.¹⁸ Their target isn’t ours, so I can’t put it in the table. If they’d like to run ours, or send the numbers, the exams are the same for everybody.

Where we agree, and where we don’t

Which number does ours agree with most? Rank the 269 players with at least 1,000 minutes this season by mPM and by each of the others still being published, and ours matches EPM’s order most closely, ahead of DARKO and LEBRON.¹⁹ EPM did best on Dunks & Threes’ own exam, so I choose to find that reassuring.

The disagreements are where it gets fun. VORP is the box score’s total, so these are the players the box score undersells by the most next to us:

The ones it oversells:

It isn’t random. Across all 269 rotation players, two things on the stat line move the gap: the more efficiently a player scores, and the more steals he gets, the more VORP likes him next to us.²⁰ Shooting a lot of threes tilts a player our way, though not beyond luck, and nothing else on the line, from usage to assists to rebounds, explains anything. All of it together explains 13% of the disagreement. The rest is what happened to the score while a player was out there, which is the part the scorekeeper can’t write down.

The modern metrics disagree with us about different players, but with the same habit. Against DARKO and LEBRON too, the more steals a player gets, the more they like him next to us, and against LEBRON, the more of his shots come from three, the more we like him.²¹ Every metric I checked this on gives steals more credit than ours does. Either they all overrate steals or we underrate them, and none of these exams can say which. Among the biggest disagreements is one I’d rather lose: DARKO has Jaylen Brown 30th this season, and we have him 137th. Ours has been low on Brown before, and it wasn’t fun then either.

Then there’s Marcus Smart, who breaks the pattern. He gets steals more often than nearly nine in ten rotation players, exactly what the box score loves, and it still has him 224th by VORP, in large part because he doesn’t shoot efficiently.²² We have him 51st by mWAR, and every metric here that watches the scoreboard has him somewhere between 58th and 103rd.

So which one should you use?

So, should you trust ours? Trust it the way you’d trust any of the good ones. In the regular season it has the smallest miss of anything here, with home court. In the playoffs, where nobody had home court, it isn’t clearly better than DARKO, RPM or RAPTOR, or than LEBRON on this season’s numbers, and none of us beats a glance at this season’s standings by much. PER and Win Shares are clearly behind. For a single season’s MVP argument, use mWAR, because playing counts. I’ll rerun the playoff exam every spring, and if one of them pulls ahead of ours, you’ll read it here first.

Figures: 2025-26 from Basketball-Reference (Win Shares, VORP, BPM, PER), Dunks & Threes (EPM, as of 2026-06-13) darko.app (DARKO, its histories too) BBall Index’s LEBRON dashboard (LEBRON) and saved ESPN season tables (RPM); RAPTOR and PREDATOR from FiveThirtyEight’s final release, and RAPTOR’s history from its data archive; older seasons’ BPM, Win Shares and PER from Basketball-Reference; the exams from metric_tests.py on this site’s harness, as of this build. Wins are on this model’s own scale and dollars at the open-market price of a win; the method explains both.

1. Dan Rosenbaum, writing on 82games.com in 2004, called it by far the most any team paid an outside stats consultant. ↩
2. The richer starting guess also leans on a player’s earlier seasons, which describes him well and goes stale fast. We checked both on the regular-season exam, which is part of the home court that comes up below. ↩
3. Each metric’s name in the table links to its creator’s write-up, or the closest thing to it; FiveThirtyEight’s is gone, so RAPTOR’s links to an archived copy. ↩
4. A level the model estimates from the games rather than assumes: an mPM of −1.77 per 100. VORP sets it at a BPM of −2.0 by convention. The method page has the arithmetic. ↩
5. I also reran the playoff exam with each number as it stood on the last day of that regular season, for the nine metrics where that’s possible; RAPTOR, PREDATOR and DARKO’s files mix in the playoffs, so they sit it out. The verdict barely changes: raw plus-minus, Win Shares per 48 and PER are still clearly behind ours, LEBRON draws level, and nothing is clearly ahead. ↩
6. This is the Dunks & Threes “metric comparison” protocol: players under 250 minutes, and rookies, get a flat replacement level. It isolates the rating from the minutes, which are taken as known. PER and Win Shares aren’t measured in points, so their replacement level is set at the same percentile as everyone else’s. ↩
7. How much to trust the box score, how many seasons to pool, how to handle garbage time: each choice was scored on this exam. ↩
8. For each metric I compare its miss with ours team by team, then swap the two metrics’ results whole seasons at a time, 10,000 times, and count how often the shuffled gap is as big as the real one. For BPM, RAPTOR, DARKO and RPM that’s between one time in a hundred and one in twenty, for LEBRON 10%, and for the rest less than one in a hundred. ↩
9. On wins, the shuffled gap matches the real one 9% of the time for DARKO, 20% for LEBRON and 7% for RPM. ↩
10. Fit each team’s result on both forecasts at once, with errors grouped by season, and those six keep a small, real weight. The rest add nothing ours doesn’t already have. ↩
11. RAPTOR’s history ends in 2022-23. Older seasons’ BPM, Win Shares and PER are matched to players by name. Each metric’s scale is fit on its own seasons and ours is refit on the same ones, so ours reads a little differently in each comparison. The luck check is the same shuffle as before, era by era. ↩
12. On end-of-regular-season numbers the shuffle gives LEBRON 0.29. ↩
13. Same shuffle as before, series by series with whole seasons swapped. Share of shuffles with a gap as big as the real one: Plain RAPM 0.31; Our starting guess 0.69; Raw plus-minus 0.01; BPM 0.10; Win Shares per 48 0.01; PER 0.02; RAPTOR 0.11; PREDATOR 0.09; DARKO 0.25; LEBRON 0.03; RPM 0.61. ↩
14. The Trade Lab’s method has the formula, and the caveat that it may reward a single star more than the evidence cleanly supports. ↩
15. Typical chance given to the champion going in, worked out through the real bracket: ours 25%; Plain RAPM 27%; Our starting guess 23%; Raw plus-minus 22%; BPM 27%; Win Shares per 48 22%; PER 23%; RAPTOR 29%; PREDATOR 28%; DARKO 23%; LEBRON 23%; RPM 34%. Over RAPTOR, PREDATOR and RPM’s seasons, ours gave the champion 26%. ↩
16. All three playoff grades use the minutes players actually logged in those playoffs, so they already know about injuries nobody saw coming. That helps every metric, not just ours, and it’s why none of this is a forecast you could have made in April. ↩
17. Among rotation players, RPM’s average is about the same on both sides of the splice, and a player’s rating carries over from 2018-19 to 2019-20 about as well as it does within either version. The saved season tables miss up to 13% of a season’s rotation minutes; those players are charged replacement level, like anyone else without a number, which costs RPM a little. ↩
18. On their version, the typical miss in points per 100 possessions is 2.48 for EPM, 2.60 for RPM and 2.71 for BPM. Their target differs from ours, so these can’t go in the table. ↩
19. Rank correlations with mPM, 2025-26: EPM 0.84; DARKO 0.82; LEBRON 0.78; BPM 0.72; Win Shares per 48 0.60; PER 0.56. Our mWAR against the totals: VORP 0.78, Win Shares 0.71. ↩
20. Each player’s rank gap regressed on seven box-score rates at once, scaled so each counts per typical difference: usage -3.3 places (p 0.41); assist rate -1.2 places (p 0.76); true shooting -11.5 places (p <0.01); share of shots from three +6.6 places (p 0.10); steal rate -7.3 places (p 0.03); block rate -5.8 places (p 0.13); defensive rebound rate -1.5 places (p 0.66). Positive means we rank him higher than VORP does. ↩
21. Same regression, our rank against each one’s: steal rate -9.7 places against DARKO and -8.4 against LEBRON; share of shots from three +10.1 against LEBRON (p 0.01) and +7.0 against DARKO (p 0.05). The box score explains 12% and 10% of those disagreements. ↩
22. His true shooting is lower than 83% of rotation players’. Win Shares has him 210th. ↩