How the forecast works

What goes into the numbers on every page here, how often they've been right, and a few things the data taught us along the way.

What goes into it

Every player gets a rating for offense and one for defense, from who was on the floor for each possession over the last two seasons. Recent games count more.
A team's strength comes from three things: its Elo rating, how it has outscored people this season (with last season as a starting point), and its players' ratings weighted by the minutes each is expected to play. So when a player gets hurt or traded, the team moves with him.
Each game's chance comes from the predicted margin, with home court. Then we play out the rest of the season 10,000 times, play-in and playoffs included, and the odds you see are how often each thing happened.
It reruns every night after the games.

How often it has been right

Every model gets graded on games it hadn't seen yet. That includes ours, and it includes 538's old forecasts, from their archive.

This season

39 games graded so far

Picked the winner: 82.1%
Probability score: 0.166
Home team always wins: 51.3%

Written down for the next slate

Nov 9 · LAL at ATL → Atlanta 53%
Nov 9 · CLE at CHA → Cleveland 59%
Nov 9 · WAS at DET → Detroit 91%
Nov 9 · MEM at TOR → Toronto 64%
Nov 9 · IND at DAL → Dallas 64%
Nov 9 · DEN at HOU → Houston 60%
Nov 9 · SAS at PHX → San Antonio 68%
Nov 9 · NOP at GSW → Golden State 68%

Past seasons

About 10,500 games from 2015-16 to 2023-24. Each season is forecast using only the seasons before it.

Probability score grades the chance, not just the pick: it's the average squared miss between the chance we gave and what happened. Lower is better. Saying 50% every game scores 0.250.

Sort by

Model	Picked the winner	Probability score
Our forecast (blend + lineup ratings)	66.2%	0.212
Our blend without lineup ratings	66.1%	0.213
Lineup-adjusted plus-minus alone	65.8%	0.214
Player plus-minus alone	65.2%	0.216
Team points for minus against	65.1%	0.217
Elo (538-style)	65.0%	0.217
Chronos-Bolt (Amazon)general-purpose	64.5%	0.220
Box-score player value	64.5%	0.221
Timestral (Mistral)general-purpose	64.4%	0.219
TimesFM 2.5 (Google)general-purpose	64.4%	0.220
t0-gamma (The Forecasting Company)general-purpose	64.3%	0.220
Prophet (Meta)general-purpose	62.8%	0.225
Home team always wins	56.8%	0.245

When we say 70%, does it happen 70% of the time?

Every past game, grouped by how sure we were about the favourite. A dot on the line means the favourite won exactly as often as we said.

Close to the line all the way up. We run a little too sure on the big favourites: games we gave about 92% went to the favourite 89.5% of the time.

With tonight's injury report

Injury reports only go back to 2018-19, so this one is graded on fewer games: 4,635 from 2020-21 to 2023-24, against our forecast on the same games.

Model	Picked the winner	Probability score
Our forecast	65.5%	0.217
Ours plus tonight's injury report	66.4%	0.214

Knowing who's out before tip-off helps. It picks a few more winners, and the chances get sharper. It isn't in the live forecast yet.

Against 538

538's pregame forecasts, from their archived data, on the same games.

Seasons	Games	538	Ours
2016-17 to 2018-19CARM-Elo	3,642	65.6% · 0.215	66.3% · 0.212
2019-20 to 2022-23RAPTOR	4,440	65.1% · 0.219	65.2% · 0.219

Against CARM-Elo we were a little better on both counts, and against RAPTOR it's a dead heat on the probability score.

The general-purpose models

Chronos, Timestral, t0-gamma, TimesFM and Prophet weren't built for basketball. Each one got a single list per team: the point margin of every game it had played, oldest first, up to its last 512. It guessed the next number. The home team's guess minus the away team's became a win chance, fit only on earlier seasons, same as ours. Four of them land just under Elo, and Prophet lands further back. That's about what a plain average of a team's last 82 games gets you, which makes sense. A list of scores can't see who's hurt or who just got traded.

What the data taught us

Five things we found while building it. A couple of them surprised us.

Adding up points per game picks the winner 59.5% of the time

Our first idea was the simple one: add up every player's points per game and see which team comes out ahead. It picks the winner 59.5% of the time. Always picking the home team gets 56.8%, so it barely beats doing nothing.

Points per game only sees one end of the floor. A team's points get split among whoever is out there, so the total mostly tells you how much a team scores. It says nothing about how much it gives up.

Plus-minus counts both ends: points scored minus points allowed while you're on the floor. Same players, same games, and it picks the winner 65.2% of the time.

Picked the winner, 2015-16 to 2023-24

Home team always wins	56.8%
Add up points per game	59.5%
Elo	65%
Plus-minus	65.2%
Our forecast	66.2%

The last 10 games will lie to you

A hot stretch feels like news. Mostly it is noise. We rated every team on its last 5, 10, 20, 41, 82 and 164 games and checked which one picked winners best.

The last 10 games got 62.6%. A full season got 64.5%. Two seasons got worse again, 62.7%, because rosters change. Statisticians call the first part regression to the mean: an extreme stretch usually drifts back toward normal.

Best of all was about a season of games with the recent ones counting more, 65.1%. That is what the forecast uses. A good stretch counts, just not as much as it feels like it should.

Picked the winner, by how many games the team rating looks back

Last 5 games	61.1%
Last 10	62.6%
Last 20	64.2%
Last 41	64.4%
Last 82	64.5%
Last 164	62.7%
Recent games count more	65.1%

Is the hot hand real? About 20% of it

When a player scores more than we expected over his last 10 games, about 23% of that extra carries into his next one. For plus-minus it is about 20%. That is from 360,480 player games.

So the hot hand exists, and it is small. Part of a hot stretch is a lasting change, like a bigger role. The rest is luck, and luck fades.

Share of a 10-game streak that carries into the next game

Points above expected	23%
Plus-minus above expected	20%

How players differ isn't how teams win

We ran a method called PCA over 7,713 player seasons. It finds the main ways players differ from each other. The two biggest are size and rebounding, and how much of the offense runs through you. Together they cover 62% of how players differ.

They explain about 3% of whether a team wins with those players on the floor. What wins sits in smaller differences: shooting efficiently, and a mix of threes and defensive rebounding. Those barely register as ways players differ, which is a big part of why the box score struggles to see them.

The five main ways players differ: share of how they differ vs share of winning

Size and rebounding: how players differ	35.7%
Size and rebounding: winning	0.3%
Scoring load: how players differ	26%
Scoring load: winning	3%
Efficient shooting: how players differ	6%
Efficient shooting: winning	11.7%
Threes and boards: how players differ	5.2%
Threes and boards: winning	12.3%

The league changed shape

Group players by how they play and the last 20 years tell their own story. Spot-up shooters went from 16% of rotation players in 2004-05 to 22% in 2025-26. Star creators went from 7% to 11%. Interior bigs fell from 11% to 7%, and playmaking guards from 17% to 13%.

Nobody told the method about the three-point boom. It found it on its own.

Share of rotation players, 2004-05 then 2025-26

Spot-up shooters, 2004-05	16%
Spot-up shooters, 2025-26	22%
Star creators, 2004-05	7%
Star creators, 2025-26	11%
Interior bigs, 2004-05	11%
Interior bigs, 2025-26	7%
Playmaking guards, 2004-05	17%
Playmaking guards, 2025-26	13%

Want the methods behind these? Our Data to AI course teaches them in five-minute lessons. Data to AI

See today's forecast

Questions

What is the probability score?

It's the Brier score: the average squared gap between the chance we gave and what happened. Being sure and right scores well, and being sure and wrong costs a lot. Lower is better. Saying 50% every game scores 0.250; ours is 0.212.

When it says 70%, does that team win 70% of the time?

Close to it. Across 10,422 past games the favourite won about as often as we said, running a little short at the top: games we gave about 92% went to the favourite 89.5% of the time.

Does the forecast know who's injured tonight?

Not yet. It knows who has been playing, from each player's minutes over his team's last three weeks, so a player who missed the last game gets no minutes. Adding the official injury report picked more winners in testing. It isn't in the live forecast yet.

538 forecasts: FiveThirtyEight data (CC BY 4.0), from the Internet Archive. Our forecast, built from play-by-play data. Not affiliated with the NBA.