For this post, I’m using my NBA prediction model to simulate the entire playoffs from start to finish. This is the same model I use during the regular season, with adjustments to reflect playoff conditions.

The model is built around Current Ability (CA), a player rating I developed. These ratings are aggregated into team strength and combined with team performance data to generate game-level predictions. During the regular season, the model achieved 73.5% win prediction accuracy.

If you want a full breakdown of how the regular season model works, you can read that here.

You can also explore the full player ratings database.

Adapting the Model for the Playoffs

The core model remains unchanged, but inputs are adjusted to reflect playoff conditions.

Two key changes:

1. Playoff context features
The model incorporates series-level context such as games played to date, elimination scenarios, and other factors that influence performance within a series.

2. Playoff rotations

Team strength is calculated using projected playoff minute distributions, with tighter rotations and a higher concentration of minutes among top players.

Projected availability is also incorporated. For players expected to return later in the playoffs, their impact is introduced in the corresponding rounds (e.g. Luka Dončić or Austin Reaves returning in the second round).

Last season, the model correctly predicted 13 of 15 playoff series.

Game-Level Simulation

The simulation runs at the game level, allowing each series to evolve dynamically based on prior outcomes.

As the playoffs progress, team context updates in real time within each simulation:

• Matchups update as teams advance

• Team strength adjusts over time

• Fatigue carries from round to round

For example, if the Sixers upset the Celtics in the first round, the next opponent faces that version of the Sixers, not their original baseline. Their ratings, form, and accumulated playoff context reflect what they achieved in that series.

This creates a fully path-dependent simulation, where outcomes are shaped not just by team strength, but by how the bracket unfolds.

By repeating this process thousands of times, the model generates probabilities for:

• Advancing through each round

• Winning the Finals

• Series outcomes (in 4, 5, 6, or 7 games)

You can view the current title odds table here.

Most Likely Bracket:

One interesting takeaway:

The Pistons and Cavaliers meet in the conference semifinals in about 75% of simulations. Cleveland wins 50.77% of those series, but Detroit reaches the conference finals more frequently overall due to a higher first-round advancement rate.

Still, this isn’t the Premier League. Teams can’t simply spend their way to a title. Instead, success depends on how well teams maximize the value of their payroll. Having recently created Current Ability (CA), a player rating metric, I figured why not use it to evaluate contract value. The idea is simple: compare a player’s current on‑court value to what their Average Annual Value (AAV) implies they should be worth.

Of course, many other factors go into contract evaluation. For example, this analysis does not account for teams paying a premium for a player’s potential. But for teams trying to win now, maximizing the value of each contract is what matters most.

Percentile Map

Displays each player’s CA percentile vs. AAV percentile.

• Above the diagonal = performing better than salary rank

• Below the diagonal = performing worse than salary rank

Best and Worst Contract Values

The Current Ability level implied by each salary tier is first estimated using a fitted AAV-to-CA curve. Each player’s CA is then compared to this expectation, and the table highlights the 15 largest positive and negative gaps.

Ultimately, maximizing payroll value is a key factor separating contenders from champions. Organizations that consistently compete at the highest level allocate their cap resources efficiently. While this framework is not intended to fully capture contract value in all contexts, it provides a useful tool for evaluating potential trade targets.

After recently completing my player rating metrics, I wanted to integrate them into my prediction model to see if it improved performance. The metric used here is Current Ability (CA), which I designed to estimate a player’s true level right now. Over the All Star break I built the 2.0 version with clear goals: better account for who is actually on the floor, separate dips in form due to injuries from true decline, and distinguish variance from true team ability.

The model uses gradient boosted decision trees trained on multi‑season game data. They are a good fit here because they capture nonlinear interactions without forcing a fixed linear relationship. The inputs include season‑level team strength, recent form, and roster strength built from minutes‑weighted CA. The model predicts margin and total points, then converts margin into a win probability using the historical margin variance from the training set. It is validated on the most recent full season to check generalization.

How the Model Works

1. Team performance
The base team signal comes from season to date performance metrics and an Elo style rating that updates after each game. This captures overall quality and accounts for strength changes over time.

2. Roster quality
CA is aggregated into team level values using minutes weighted projections. The model produces two views of roster strength:

• A healthy baseline that reflects expected team CA when the roster is intact.

• An availability adjusted version that reflects who is actually expected to play based on injury information.

The model tracks drops in available CA and uses that to decide how much trust to place in recent form. When key players return, it does not just treat that as a short term boost. It also reduces the weight of recent poor form that was driven by those absences. This lets the model recognize when a decline is caused by missing talent rather than a true drop in team quality.

3. Trend and expectation adjustment
The model does not simply look at wins and losses. It evaluates how teams perform relative to expectations based on their roster ratings and opponent strength. That means a team can be treated as improving even while losing, if the underlying performance is strong. For example, if the Wizards drop three straight games but play three strong opponents close, the model treats that as a positive signal. At the same time, it avoids over‑rewarding wins against teams that look strong at the season level but are missing multiple key players. This helps it capture true shifts in team performance without overreacting to misleading outcomes.

Power Rankings

In addition to individual game predictions, the model can also be used to generate power rankings. Short‑term form features are held constant, and only players with season‑ending injuries are excluded. Each team is then simulated against every other team home and away, and the projected margins are averaged. These outputs can be used to rank the league in a way that emphasizes underlying team strength, closer to a title‑odds style view of the league at any point in time.

This post introduces my roster construction model, built around three projects I’ve been working on:

• Player ratings: Season Rating (SR), Current Ability (CA), and Potential Ability (PA), used to evaluate individual players (learn more here).

• Player archetypes: labels that classify a player’s role on both offense and defense.

• Lineup synergy scores: a way to identify archetype combinations that consistently overperform after adjusting for player ability.

The purpose of the roster construction model is to identify the archetypes that would strengthen a team the most, then search within those archetypes for player targets. How those targets are ranked depends on the phase the team is in. For a team trying to compete now, the model prioritizes CA and estimates how much each player would increase the team’s projected Net Rating. For a rebuilding team, PA matters more, because the priority is identifying who is most likely to develop into an elite player.

Archetypes

To classify players’ on-court roles on both sides of the floor, I built offensive and defensive archetypes using k-means clustering. The approach and archetype labels are inspired by BBall Index and CraftedNBA.

These archetypes are important for the roster construction model because they let the problem be framed as role fit, not just “add the best player available.” Two players with similar ratings can still have very different impact on a specific team, because one might fill a role the roster is missing or bring a skill set that better complements the players already in place.

The maps below visualize the archetype clusters in 2D.

You can view the complete player ratings and archetypes database in the NBA Insights Ratings Database.

Lineup Synergy

By defining lineups through their five-archetype combinations, I developed synergy scores that measure how much a given archetype combination performs better or worse than expected based on the CA of the players on the floor.

Synergy scores are calculated using the same stint-based regression approach used for RAPM, but applied to archetype combinations instead of individual players. In RAPM, lineup overlap is used to quantify individual impact while adjusting for who a player shares the floor with and who they face. The goal here is to figure out how well different archetypes complement each other. However, stint results are still heavily influenced by ability, so if one lineup is clearly stronger than the other, it will usually win that stretch of possessions even if the archetypes don’t fit together particularly well.

With that in mind, the model first luck-adjusts for free throw and three-point shooting variance, then predicts each stint’s net rating using only (i) the CA difference between the two lineups and (ii) a home-court adjustment. The gap between the observed result and the model’s prediction is then grouped by archetype combination and aggregated across thousands of possessions, revealing which combinations consistently overperform relative to the quality of the players on the floor.

For example, an archetype combination consisting of 2 Spot Up Shooters, 1 Playmaker, 1 Roll Man, and 1 Shot Creator recorded a synergy score of +4.8 over 13,345 possessions from 2018 to 2026, meaning it has, on average, performed 4.8 points per 100 possessions better than what the baseline model expected. To put these scores in context, the chart below shows the distribution of lineup synergy scores for combinations that meet the minimum possessions threshold.

Below are a few of the highest- and lowest-scoring archetype combinations by lineup synergy:

Highest

• 2 Spot Up Shooters, 1 Playmaker, 1 Roll Man, 1 Versatile Big: +9.9

• 1 Playmaker, 2 Roll Men, 1 Movement Shooter, 1 Shot Creator: +9.4

• 1 Playmaker, 1 Movement Shooter, 2 Shot Creators, 1 Connector (Big): +8.0

What these lineups have in common is balance. They combine playmaking, shooting, and players who can create their own scoring opportunities, while still keeping enough size on the floor to hold up physically, finish possessions with rebounding, protect the rim, and avoid getting punished in the paint.

Lowest

• 2 Spot Up Shooters, 1 Playmaker, 1 Shot Creator, 1 Connector (Small): -9.8

• 1 Spot Up Shooter, 1 Movement Shooter, 1 Secondary Ball Handler, 1 Shot Creator, 1 Connector (Small): -8.3

• 1 Roll Man, 2 Movement Shooters, 1 Shot Creator, 1 Connector (Big): -7.1

Small-ball shows up disproportionately in the worst results, which you can see in the first two examples above. Five of the bottom ten lineups are small-ball groups, compared to zero of the top ten. That is consistent with the idea that small lineups no longer create the same spacing advantage they once did. As more bigs can shoot, you can keep the floor spaced without giving up size.

Notice that 1 Playmaker, 2 Roll Men, 1 Movement Shooter, 1 Shot Creator (+9.4) and 1 Roll Man, 2 Movement Shooters, 1 Shot Creator, 1 Connector (Big) (-7.1) are structurally very similar. The key difference is the presence of a playmaker in the top version. Movement shooters create open looks by navigating screens, and without a playmaker to consistently capitalize on those openings, that extra movement shooting can be harder to convert into real advantage.

These results suggest that there is real value in being able to consistently play rotations that produce strong archetype combinations. A team might not be able to run one specific five-man group every night, but a well-built roster can repeatedly put lineups on the floor that fit together at the archetype level, even when players are missing. That idea is what the roster construction model is built to optimize.

Roster Construction

I designed the roster construction model to analyze a team’s roster composition and identify additions that unlock the strongest archetype combinations, whether working with the current roster or exploring hypothetical trade scenarios. The model begins by mapping the distribution of offensive archetypes of a given roster, then searches for which archetype, or combination of archetypes, would produce the largest total synergy score (the sum across eligible combinations) when inserted into specific roster slots.

By optimizing for total synergy score, the model rewards additions that expand the set of archetype combinations with strong synergy scores that a team can access across its rotation. The highest-rated moves are the ones that unlock more consistently positive combinations, increasing the number of viable five-man groups that fit together at the archetype level over the course of a season.

The model has two core components: an archetype search and a trade evaluator. The archetype search is an open-ended exploration of roster needs, identifying the archetype additions that maximize total synergy score across lineup combinations that clear both the minimum possessions requirement and the synergy score threshold. The model then returns the archetypes and combinations that best increase team synergy, along with player targets within the recommended archetypes. The trade evaluator is designed for testing hypothetical trade scenarios. After identifying a shortlist of targets through the archetype search, specific additions and removals can be tested to estimate the net change in total synergy score, projected Net Rating, and expected wins.

What this looks like in practice

Using the Minnesota Timberwolves as an example, the model projects a team Net Rating of +4.4 and an expected win total of 51 games. It arrives at that projection by aggregating the roster’s CA based on the share of possessions each player is expected to play, then mapping roster CA to historical team Net Rating and Net Rating to expected wins. That baseline lines up closely with Minnesota’s performance to date, with a current Net Rating of +4.5.

The archetype search identified a Playmaker as the best fit by total synergy score, which aligns with Minnesota’s need for additional ball-handling and creation. Using the trade evaluator, I then tested a hypothetical scenario where Minnesota trades Rob Dillingham and Terrence Shannon Jr., plus additional assets, for Tre Jones. The model projects that this swap would increase the team’s Net Rating from +4.4 to +5.3, raising expected wins from 51 to 53 over a full season. Jones also fits the team’s timeline. He is only two years older than Anthony Edwards, and his PA, shown in the table below, suggests he still has room to develop.

This trade example is just one way to use the model. It can support teams at different stages, from identifying archetypes to target while building around a young core to evaluating specific win-now trades.

Predicting the Entire NBA Playoffs Using Machine Learning

(Perfect 8-for-8 Start) Predicting the NBA Playoffs Using Machine Learning

Predicting the NBA Playoffs Using Machine Learning: Conference Finals Projections and Title Odds

Conference Finals Recap

For those new to the series, at the beginning of the 2025 NBA Playoffs, I published a full postseason forecast and have followed it with updated projections after each round. The forecast is based on a machine learning model I built, which runs 10,000 simulations per update and dynamically adjusts for each series outcome by factoring in rest advantages, momentum, team Elo, and matchup context to reflect the evolving postseason landscape.

The predictions have performed well, correctly forecasting 12 out of 14 series so far. The entire first round was predicted perfectly, with the only misses coming in the Conference Semifinals: Cavaliers vs. Pacers and Celtics vs. Knicks—both notable upsets that were also influenced by injuries.

As outlined in my previous post, I provided probabilities for each Conference Finals series, not only predicting the winner, but also the number of games. The model correctly forecasted Pacers in 6 and Thunder in 5.

Model Picks vs Betting Odds

Betting lines finished 9–5, while my model went 12–2. Notably, all six series with odds within the ±200 range were predicted correctly.

Finals Projections

After 10,000 simulations, my model estimates that the Thunder are most likely to win the series in 5 games (24.2%). I think this prediction makes sense, as the Pacers have a good chance to take one of Games 3 or 4 at home.

Throughout the playoffs, the Pacers have leveraged their depth and pace effectively. But unlike their earlier matchups, the Thunder have a deeper bench, the best defense in the league, and thrive playing at a high pace. The Nuggets proved that the Thunder struggle when forced into a slower, half-court game, something the Pacers are unlikely to replicate. Even if Indiana try to slow things down, it would come at the cost of their own strengths, making it difficult for them to gain an edge in this matchup.

I've really enjoyed covering and predicting the postseason this year. If the Thunder close it out, my model will finish with a 13–2 record on series predictions, a strong result. I'll continue posting game-level predictions and updated series odds throughout the Finals. Thanks for reading, and if you enjoyed this post, consider subscribing so you don’t miss the next one.

OKC vs. IND Game 1 Projection & Series Odds – 2025 NBA Finals

Subscribe now

Predicting the Entire NBA Playoffs Using Machine Learning

(Perfect 8-for-8 Start) Predicting the NBA Playoffs Using Machine Learning

For those new to the series, at the beginning of the 2025 NBA Playoffs, I published a full postseason forecast using a machine learning model I built. Across 10,000 playoff simulations, the model dynamically adjusted for each series outcome, updating rest advantages, momentum, team Elo, and matchup context to reflect the evolving postseason landscape.

The first round went perfectly, my model correctly predicted all eight series, and I entered the conference semifinals confident that the second round would follow suit. Unfortunately, that didn’t last. Two major upsets followed, with the Knicks and Pacers defying expectations to advance.

As shared in my last post, these were my model’s projections heading into the second round:

West

70% OKC vs. DEN 30% ✅

52% MIN vs. GSW 48% ✅

East

67% CLE vs. IND 33% ❌

81% BOS vs. NYK 19% ❌

One interesting note: when I first ran my model at the start of the Playoffs, it projected a Timberwolves–Warriors matchup in the conference semifinals and had Golden State advancing. But after the first round concluded and I reran the model, the prediction flipped, this time giving the Timberwolves a slight edge.

That shift says a lot about how the model accounts for both performance and fatigue. Minnesota was rewarded for the dominant way they handled the Lakers, while the Warriors were penalized for needing seven games to get past the Rockets. Of course, things might have played out differently if Curry had been fully healthy, but it’s a good example of how postseason dynamics can swing projections, even in close matchups.

Conference Finals Forecast (10,000 Simulations):

🏆 Championship Odds:

Oklahoma City Thunder: 51.4%

Indiana Pacers: 17.8%

Minnesota Timberwolves: 17.3%

New York Knicks: 13.5%

After including the Finals in the simulation, the Thunder have unsurprisingly emerged as the clear title favorites, winning in 51.4% of all runs. The Pacers and Timberwolves have nearly identical championship probabilities (17.8% and 17.3%, respectively), but achieved through different paths. Indiana reached the Finals in 55.6% of simulations, while Minnesota did so in only 32.7%, implying that the Timberwolves have a much higher conditional win rate once they reach the Finals. This suggests Minnesota is stronger head-to-head against Eastern Conference opponents, even though they are less likely to get there. Based on overall title equity, the implied power ranking is: Thunder > Timberwolves > Pacers > Knicks. Regardless of outcome, my model favors the Western Conference representative in the Finals.

Thanks for reading, if you enjoyed this post, consider subscribing so you don’t miss the next one.

Subscribe now

At the beginning of the 2025 NBA Playoffs, I published a full postseason forecast using a machine learning model I built. Across 10,000 playoff simulations, the model dynamically adjusted for each series outcome, updating rest advantages, momentum, team Elo, and matchup context to reflect the evolving postseason landscape.

Simulated Bracket at the Start of the Playoffs

I wasn’t sure how well my playoff model would perform. Unlike my regular-season model, which I’ve been able to tweak and refine continuously throughout the year, I built this one in isolation, with no opportunity to adjust it as results came in.

One of my biggest uncertainties was the decision to train the model exclusively on playoff data, rather than a combination of playoff and regular season games. The tradeoff was clear: by using only playoff games, I had to go much further back in time to gather enough data, which naturally introduces the risk that older games reflect a different version of the sport. On the other hand, the regular season can feel like an entirely different game, played at a different pace, with different rotations, and often different levels of intensity.

Despite that uncertainty, the results so far have been very encouraging. The model correctly predicted all 8 first-round series, including four that oddsmakers saw as toss-ups. Notably, it favored both the Timberwolves and Nuggets, even though they entered their series as betting underdogs.

Subscribe now

Updated Forecasts and First-Round Recap

I've included updated forecasts for the Conference Semifinals and Finals that take into account results from the first round. Below that is a brief recap of all eight first-round series from my perspective.

West

70% OKC vs. DEN 30%

52% MIN vs. GSW 48%

East

67% CLE vs. IND 33%

81% BOS vs. NYK 19%

Three of the four series still align with my original projections; however, my model now gives the edge to the Timberwolves. This shift is likely due to how the first-round results played out. My model had projected the Warriors to beat the Rockets in fewer games and expected a closer series between the Timberwolves and Lakers. It also viewed both the Wolves and Lakers as favorable matchups for Golden State. With the outcomes reversed, it makes sense that the model now favors Minnesota. They beat the Lakers in dominant fashion and will enter the next round with a rest advantage.

Hopefully it is wrong and my original bracket stays alive…

🏆 Championship Odds

Boston Celtics: 27.6% — ↑ 10.7%

Oklahoma City Thunder: 22.8% — ↑ 8.0%

Cleveland Cavaliers: 14.5% — ↑ 3.2%

Minnesota Timberwolves: 11.9% — ↑ 4.1%

Golden State Warriors: 8.6% — ↑ 0.7%

Denver Nuggets: 7.0% — ↓ 1.6%

Indiana Pacers: 4.7% — ↑ 1.5%

New York Knicks: 2.9% — ↓ 1.4%

[OKC] Thunder vs. [MEM] Grizzlies
Model Prediction: Thunder 89% – Grizzlies 11%
Result: Thunder in 4 ✅

After two dominant home performances from the Thunder, the Grizzlies came out with everything they had and built an impressive 26-point halftime lead in Game 3. However, an injury to Ja Morant shifted the momentum in the Thunder’s favor as they stormed back and eventually closed out the series 4-0.

[CLE] Cavaliers vs. [MIA] Heat
Model Prediction: Cavaliers 83% – Heat 17%
Result: Cavs in 4 ✅

Not much to say about this series, the Cavs decimated the Heat, outscoring them by a historic 122 points over four games. Donovan Mitchell had a strong start to the playoffs, averaging 23.8 points per game, with notable contributions from Ty Jerome, Evan Mobley, and Jarrett Allen.

[HOU] Rockets vs. [GSW] Warriors
Model Prediction: Rockets 43% – Warriors 57%
Result: Warriors in 7 ✅

This series quickly turned into a low-scoring but highly entertaining slugfest, defined by four major momentum swings. After Game 1, it looked like the gap in experience and offensive firepower would be too much for the Rockets to overcome. But after Jimmy Butler went down with an injury in Game 2, the Rockets responded with a dominant performance. Jalen Green led the way, temporarily silencing the doubters with a 38-point masterclass—before quickly proving them right again.

With Butler still out in Game 3, it seemed like the Rockets might carry that momentum forward and take control of the series. Instead, the Warriors took both Games 3 and 4 despite Butler’s absence. Gary Payton II delivered an explosive fourth quarter in Game 3, and Brandin Podziemski followed with a huge 26-point performance in Game 4, helping the Warriors build what appeared to be an insurmountable 3–1 lead.

Out of nowhere, Fred VanVleet caught fire, hitting an incredible 18 of 27 three-pointers over Games 4 through 6. He also brought the composure that the young, inexperienced Rockets needed, helping lead them to two convincing wins and forcing a Game 7. Alongside him, Amen Thompson stepped up with impressive two-way play, further showcasing his potential and giving Rockets fans a glimpse of a very bright future.

In the end, the Warriors’ experience prevailed. Buddy Hield scored a playoff career-high 33 points to keep Golden State in control, and Steph Curry closed the door with a dominant fourth quarter to seal the series.

[BOS] Celtics vs. [ORL] Magic
Model Prediction: Celtics 88% – Magic 12%
Result: Celtics in 5 ✅

The banged-up Celtics handled the Magic in a very physical series that saw Jayson Tatum injure his wrist and Jrue Holiday strain his hamstring. Jaylen Brown stepped up with a crucial 36-point performance in Game 2 during Tatum’s absence, playing at an elite level despite dealing with a knee issue. I agree with my model’s projection that the Celtics are the title favorites. I don’t believe the Cavaliers or Thunder pose as much of a threat as their regular-season records suggest. If Boston stays healthy, I’m confident they can repeat as champions.

Franz Wagner and Paolo Banchero had a fantastic series. If the Magic can build the right roster around them, I wouldn’t be surprised to see them make a deep playoff run in the next few years.

[LAL] Lakers vs. [MIN] Timberwolves
Model Prediction: Lakers 46% – Timberwolves 54%
Result: Timberwolves in 5 ✅

Ant-Man, Batman, Superman… Wolves in five. This series exposed some serious holes in the Lakers' roster while also highlighting the Timberwolves as a legitimate title contender after a relatively quiet regular season. Nobody disputed that the Lakers lacked a true center heading into the postseason, but after securing the third seed and playing solid basketball down the stretch, it didn’t seem to matter much. Going into the series, any argument in favor of the Timberwolves was often met with, “No way Luka and LeBron lose in the first round.”

But with Anthony Edwards playing at an MVP level, a healthy Julius Randle, and a strong supporting cast, the Timberwolves closed it out in five. I don’t think Lakers fans should be too worried. Austin Reaves went ice cold, and even so, the series was closer than the 4–1 result suggests. Make a few smart offseason moves, give Luka someone he can lob it to, and I think the Lakers will be back stronger next year.

[NYK] Knicks vs. [DET] Pistons
Model Prediction: Knicks 56% – Pistons 44%
Result: Knicks in 6 ✅

I was surprised that my model only gave the Knicks a 56% chance of winning this series, as I expected there to be a clearer gap between the two teams. Cade Cunningham took his game to another level this regular season, and that carried over into the playoffs. Unfortunately for the Pistons, Jalen Brunson and Karl-Anthony Towns showed their experience down the stretch, closing out several tightly contested games.

[DEN] Nuggets vs. [LAC] Clippers
Model Prediction: Nuggets 51% – Clippers 49%
Result: Nuggets in 7 ✅

“If you like basketball, like a really true fan of basketball — not like fame basketball, like really, like details stuff — I think this is the games that you should watch.”
– Nikola Jokić

In my opinion, this was the best first-round series by far. It was high-level basketball at its finest: physical, tactical, and competitive from start to finish, with two elite duos going head to head in a true chess match. Every game felt meaningful, with momentum constantly shifting and both teams making tough in-game adjustments.

My model projected this series to be about as close as it gets, so a Game 7 outcome came as no surprise. The Nuggets took control in the final game, with six different players scoring in double figures. Unfortunately for Denver, their reward for surviving that battle is a second-round matchup against a rested and confident Thunder team just two days later.

As for the Clippers, it is hard not to start asking difficult questions. They have now had four first-round exits, one trip to the conference semifinals, and one conference finals appearance over the last seven seasons. James Harden’s Game 7 struggles continued, turning in another underwhelming performance when it mattered most.

And then there is the Shai Gilgeous-Alexander trade. With SGA now an MVP candidate leading the number one seed in the West, that trade is starting to look like one of the worst in modern NBA history.

[IND] Pacers vs. [MIL] Bucks
Model Prediction: Pacers 55% – Bucks 45%
Result: Pacers in 5 ✅

Tyrese Haliburton looked as sharp as ever, leading the Pacers to a 4–1 series win over the Bucks in what turned out to be a statement performance for Indiana. Giannis Antetokounmpo did everything he could to keep Milwaukee alive, including a 30-point, 20-rebound, 13-assist performance in Game 5, only to come up short once again. Damian Lillard had just returned after missing over a month due to deep vein thrombosis, only to suffer a torn Achilles early in Game 4.

It’s safe to say this just wasn’t the Bucks’ year. With Giannis likely to move on this summer, the focus shifts to what Milwaukee can do with the players and picks they receive in return.

Thanks for reading. I’ll have more playoff content coming soon, including game analysis and updated predictions. If you enjoyed this post, consider subscribing so you don’t miss the next one.

Subscribe now

Predicting the NBA Playoffs is always a challenging task—especially in a year like this, where the matchups are tight and there's no clear favorite. While my regular-season model has achieved over 80% accuracy during the period I’ve tracked it, I chose not to use it here. That model relies on features relevant to the regular season that don’t translate well to the postseason.

Instead, I built a custom playoff model trained exclusively on historical playoff data, using features tailored specifically to the postseason, including:

• Current series score

• Whether a team can be eliminated or advance

• Result of the previous game

• Total games played (to account for rest or fatigue between series)

• And more

To simulate the full postseason accurately, I created a system that dynamically updates these features as the simulation progresses. For example, in a simulated run where the Thunder sweep the Grizzlies, the model adjusts for their updated Elo, win streak, rest advantage, and more when facing their next opponent (e.g., the Nuggets or Clippers).

Pre-Playoff Elo Ratings

Because Elo plays a key role in the model’s predictions, here are the Elo ratings heading into the playoffs:

Simulation Results

After 10,000 simulations, below is the predicted bracket.

The numbers behind the first round predictions are as follows:

West

89% OKC vs. MEM 11%

51% DEN vs. LAC 49%

54% MIN vs. LAL 46%

57% GSW vs. HOU 43%

East

83% CLE vs. MIA 17%

88% BOS vs. ORL 12%

56% NYK vs. DET 44%

55% IND vs. MIL 45%

🏆 Championship Odds

Boston Celtics: 16.9%

Oklahoma City Thunder: 14.8%

Cleveland Cavaliers: 11.3%

Denver Nuggets: 8.6%

Los Angeles Clippers: 8.4%

Golden State Warriors: 7.9%

Minnesota Timberwolves: 7.8%

Los Angeles Lakers: 7.5%

New York Knicks: 4.3%

Houston Rockets: 3.9%

Indiana Pacers: 3.2%

Milwaukee Bucks: 2.3%

Detroit Pistons: 1.4%

Memphis Grizzlies: 0.9%

Orlando Magic: 0.5%

Miami Heat: 0.3%

Non-Elo Simulation:

Curious about the influence of Elo on the playoffs, I reran the simulations with Elo completely removed. This version relied purely on playoff-specific features and advanced player/team stats.

The result? A surprising highlight: the Clippers emerged as a dark horse contender for the Finals.

• They still only beat the Nuggets in 49% of simulations

• However, in the simulations where they advanced to the second round, they reached the Conference Finals 58% of the time—compared to the Elo-based model, where the Thunder were heavily favored in that matchup

• Once in the Conference Finals, they advanced to the Finals in 61% of those runs

However, the model shows that if the Clippers do manage to reach the Finals, their chances drop again—winning the championship only 42% of the time overall. If they face Boston specifically, their win probability falls to just 37%, making the Celtics a particularly tough matchup in this hypothetical run.

Thank you for reading! I’ll be posting daily match predictions throughout the playoffs starting tomorrow — subscribe to stay updated.

Subscribe now

The model forecasts game outcomes by combining three primary signals: team strength, roster-level performance aggregates, and short-term form. Team ratings evolve continuously throughout the season and partially carry over between years to balance recency and stability, capturing how teams rise or decline over time.

Player statistics, including per-100 and advanced metrics, are minutes-weighted into team features using a rotation-based aggregation process. Current team profiles reflect expected availability based on recent injury information. The model then generates a win probability and projected final score for each matchup.

Modeling Approach

A gradient-boosted classifier (XGBoost) estimates the probability of a team winning, while two gradient-boosted regressors project team points scored and allowed. Hyperparameters are tuned with Bayesian optimization to identify the most effective parameter ranges.

The model is trained on a multi-season dataset that includes all available games through the current year. During training, it randomly samples across seasons and game contexts to capture a broad distribution of matchups. To improve robustness, an ensemble of models is trained with different random seeds, and their outputs are averaged to create a more stable and consistent prediction for each game.

Evaluation

The classifier is evaluated using Accuracy and AUC, and the regressors using MAE and RMSE. Across multiple random seeds, the win classifier achieves an average accuracy of 76.1% ± 1.27% and an AUC of 0.835 ± 0.011.

Together, these elements create a balanced forecasting system that adapts as the season develops. The goal is not to predict every upset, but to estimate realistic win probabilities based on team quality, roster dynamics, and recent form.

In last year’s postseason, the playoff version of the model finished with a 13–2 series prediction record, outperforming betting markets that went 10–5. See the full playoff recap in the post below.

Subscribe now