How Our NBA Title Simulations Work: 10,000 Seasons and a Predicted Bracket

"Who wins tonight?" and "who wins the title?" look like the same kind of question, but they are not. The first is one game between two known teams. The second depends on up to 1,230 regular-season games, a play-in, four rounds of best-of-seven and matchups that nobody can name yet.

This article explains how our title simulations answer the second question, how 10,000 simulated seasons become one predicted playoff bracket, and why this needs a different model from our daily game predictions.

Two questions, two models

Our daily NBA predictions come from neural networks that look at each scheduled game in detail: the expected starting five, the injury report, rest and recent form of the two teams playing that night. That detail only exists a few hours before tip-off. It does not exist for a second-round series next month between two teams we do not know yet.

A title forecast has to give a win probability for every possible matchup, including hypothetical ones. For that we use team strength ratings: the same Elo system behind our power rankings, built from NBA results since 1997. Any two teams, at any moment, have a rating gap, and the rating gap gives a win probability.

Then we play the season 10,000 times

1. Keep what already happened. Games already played are fixed. A playoff series that stands at 3-1 is resumed from 3-1.
2. Decide every remaining game by chance, weighted by strength. The weights come from the two ratings, home court and whether a team is on the second night of a back-to-back. A team with a 65% chance wins that game in about 65% of the runs. Upsets are part of basketball, so the simulation keeps them.
3. Play to the end. Finish the regular season, set the standings, play the play-in, then every best-of-7 series in the real 2-2-1-1-1 home-court format until there is a champion. Record the result and start again, 10,000 times.

The rest is counting. A team that wins 1,200 of the 10,000 simulated seasons has a 12% title chance. A team that reaches the conference finals in 3,100 of them has a 31% chance of getting that far. The same runs give the title odds, the chance of reaching each round and the bracket below.

The predicted bracket, and what "OKC in 6" means

Every simulated season produces a complete bracket, so we can show the playoff bracket we expect, series by series. For each slot we show:

• The most likely matchup. Before the field is set, a slot can host different pairings. We show the one that came up most often, and how often.
• Who wins the series. Each team's chance of taking the series, if that matchup happens.
• The most likely score. "OKC in 6 · 4-2 · 36%" means that in the runs where this matchup happens, Oklahoma City winning 4-2 was the most common result, in 36% of them. Tap a series to see every score from 4-0 to 4-3 for both teams.

Once a real series starts, its actual score is fixed and only the rest is simulated.

Why the standings are not enough

The path matters as much as the record. Two teams with the same record can face very different playoff roads: one gets a weak half of the bracket, the other has to beat two contenders to reach the conference finals. The simulation plays both roads thousands of times, so the best record and the best title odds often belong to different teams.

It is also why a team's odds can rise on a night it does not play: when a strong rival loses on its side of the bracket, its road gets easier.

What a 30% title chance means

Play the rest of the season from today many times, and a team with a 30% chance wins the title in about 3 of every 10. It does not mean the team will not win, and it does not make the forecast wrong if it does not. Surprise champions happen. An honest forecast gives them small but real chances in advance.

A probability is trustworthy when it is calibrated: things we give 40% should happen about 40% of the time. You can check this for every model on the model transparency page.

Our track record

We tested the engine on 9 past NBA seasons (2016-2025). Each season was forecast again from several points in the year, with a model trained only on data through 2015. Across all 50 forecasts, the eventual champion was our top pick 36% of the time and in our top 5 80% of the time:

For comparison, picking a champion at random from 30 teams works about 3% of the time. Forecasts are sharpest around midseason, when there are enough games to rate teams and late-season resting has not started yet. The table comes straight from the backtest files and changes only when the engine is retrained and tested again.

Updated after every night of games

After each night of games we add the new results, update the ratings and run all 10,000 seasons again. The numbers you see in May already include everything that happened since October.

Limits worth knowing

• Home court and back-to-backs are included. A team on the second night of a back-to-back is rated slightly lower for that game, and the 2-2-1-1-1 format gives the higher seed its home-court edge.
• Played games are never simulated. Results and finished series are fixed. Only what is still undecided is a probability.
• 10,000 runs keep the noise small. The random error on a title chance is under half a percentage point, so day-to-day moves come from new results, not from chance.
• Ratings only learn from results. A mid-playoff injury shows up only once it changes results, and a team that rests its stars in April looks weaker than it is.