In Game 4, the New York Knicks pulled off the largest comeback in NBA Finals history — roaring back to shock the San Antonio Spurs after being down 29 points in the second quarter.

In a related story, Karl-Anthony Towns barely played in the first half due to foul trouble, and then played 18 minutes in the second half.

Towns is now +48 in the NBA Finals when he’s on the court, while the Knicks are -40 when he’s off, prompting this question:

Is he the very best center in the NBA to go against Victor Wembanyama?

In other words, is KAT the Wemby antidote?

Let’s answer those questions, and look at (a) how Towns is doing it and (b) which other centers are best equipped to give Wembanyama trouble.

Read more

The NBA draft is just two weeks away, and it’s about time for teams to make their decisions — on what is considered an especially strong bunch of prospects.

Here is how I would make the top 14 picks:

Read more

The 2026 NBA draft class is a strong one, especially when you look metrics based on box-score stats, such as BPM.

But what does lineup data tell us about this year’s class?

Who are the standouts? Whose rating seems like a big red flag?

Let’s dive in.

Read more

In an alternate timeline, Jalen Brunson is still a Dallas Maverick.

“I really did want to stay in Dallas,” he would later say.

Instead, Brunson is headed to the NBA Finals with the New York Knicks.

From 2019 to 2022, I worked in the Mavs’ front office, providing a unique vantage point on the blunders that cost Dallas a future All-NBA point guard — how the franchise, then led by majority owner Mark Cuban1 and general manager Nico Harrison, fumbled Brunson away.

How exactly did we get to the point where the infamous Luka Dončić trade is in contention to be only the Mavs’ second-most disastrous decision of the past four years?

This is my version of that story — the story of how Jalen Brunson put himself in the position to become one of the greatest Knicks ever while making the Mavs’ blunders more and more painful.

Read more

Last year, Victor Wembanyama visited the Shaolin Temple in China for extensive training.

He was coming off a sophomore NBA season in which he had played just 46 games because of minor injuries followed by a more serious diagnosis of deep vein thrombosis, a blood clot in his right shoulder that ended his season early.

The journey bore fruit. He went on to have his best season of his young career, improving at a pace we rarely see among players already so impactful. This past regular season, Wemby surpassed Nikola Jokić in NBA impact, as measured by xRAPM — a metric that had been dominated by the Serbian for half a decade before the rise of Shai Gilgeous-Alexander and Wemby.

What Wemby did there, and why it worked, is what this piece is about.

As someone who has practiced meditation for more than a decade, I can address these questions with some specificity:

• What are the techniques that helped Wemby’s rapid ascent to the top of the NBA?

• What are the benefits?

• How can other NBA players — or any of us — take the same strides toward peak performance?

Read more

The NBA puts each prospect through a series of tests and measurements at its annual draft combine:

• The tests measure speed, quickness, agility and leaping ability.

• The measurements include height, weight, wingspan, standing reach, hand length, and hand width.

But what does it all mean? It’s hard to know intuitively which results matter, and which do not.

To solve that, we can plug the combine results into the NBA draft model and let machine learning decide which numbers have the strongest predictive value.

This will also tell us which players’ draft stock should be most influenced by these numbers.

Let’s break down the results to talk about what matters and what doesn’t from the NBA draft combine — according to my draft model.

Read more

A number of NBA teams find themselves in an awkward situation after Sunday’s draft lottery — they have a potential mismatch between what they want and who will be available.

As a result, we might see an unusual flurry of draft-day trades at the top of the lottery, with teams moving up — and down.

Here’s how I think things could play out, with three big trades right at the top of the draft:

Read more

AJ Dybantsa is the consensus No. 1 pick in the NBA draft, or at worst top three.

He’s got a unique combination of size and scoring ability, and the team that drafts him is likely to thrust him into a high-usage role almost immediately.

But my statistical draft model has doubts. While it agrees with the consensus that this is a very strong draft class, Dybantsa is seen as a very risky pick: He’s about twice as likely as Cam Boozer to become a bust.

Let’s look at some of Dybantsa’s stats — and historical comps — to understand why the model is skeptical.

What is a BUST?

And how many former high draft picks qualify?

Read more

Let’s go to the game film — and of course, advanced stats — to evaluate the top prospects in this loaded NBA draft class.

Here’s my ranking of the top 30 players available:

Read more

• One that uses lineup data (aka plus-minus), which is how a team performs with a player or set of players on the floor.

• Another that uses individual player stats, including those found in the box score.

Lineup data can, in theory, account for every action that make a difference on a basketball court. But it can be extremely noisy and requires a larger sample size.

Individual stats — points, rebounds, assists, etc. — are less susceptible to noise and stabilize a lot faster.

If we want to estimate a player’s impact as accurately as possible —- even if (and especially when) the sample size is limited —- we need to make use of all the available data.

Creating a box-score metric, similar to Basketball Reference’s BPM, is a crucial step in the process.

This article will show you how.

What we’re trying to find

With advanced metrics, we want to come up with a rating for each player on both ends of the floor.

Metrics using individual data — we’ll call them box-score metrics for now — do one thing at their core: They multiply individual player stats with weights and then add them up, leading to two numbers per player, denoting the player’s impact on offense and defense.1

To avoid using subjective estimates — which would prove suboptimal in generating accurate metrics — we need techniques for generating these weights objectively.

We can thank Dan Rosenbaum for coming up with the following method:

• Each player serves as one observation in our dataset.

• The variables are the different individual stats.

• The target variable is the player’s impact as measured by a metric that doesn’t contain any individual data.

Rosenbaum dubbed this approach “statistical plus/minus” in 2004.

The approach we present today follows in Dan’s footsteps, but with more modern techniques.

Which individual stats do we put in the model?

While the NBA box score provides the starting point, we can incorporate other stats as well.

Thanks to the NBA, we have new stats being recorded and published these days — deflections, screen assists, box-outs and many more. Some of these are derived from optical tracking, which didn’t exist when Rosenbaum originated SPM.

Beyond that, we can extract stats from NBA play-by-play data. For instance, I include kicked balls and goaltends, and I split blocked shots into two categories, depending on whether the offense or defense got the rebound.

For those who have access to raw camera data — as I did at the Dallas Mavericks — the possibilities seem almost endless: One could potentially extract how often a player misses a closeout or trails behind on an opponent fast break.

An important note: As we’ll see later, it makes sense to emphasize extracting defensive statistics more than offensive stats.

As stats come in different flavors — totals, per game, per 36 minutes, per 100 possessions, rate stats, etc. — one will have to make a decision on what to use here. My personal preference is to use the stats generated with this technique:

What’s the target variable?

As a target variable, we need a metric that a) captures the player’s impact, but also b) doesn’t include any box-score info while doing so.

That’s why we use pure lineup metrics for this step.

Lineup metrics also come in many different variations, including raw plus-minus, on-off, APM and RAPM.

Of the above, RAPM performs best at predicting out-of-sample data, so that’ll be our go-to target variable. The scale of the resulting metric will consequently mirror RAPM: Impact is measured in “influence on team points per 100 possessions,” with values around 3 and above denoting the player is a superstar.

In a way, we are treating RAPM as the ground truth. As you’ll see later, we will try to account for some of its deficiencies.

The general practice has been to use multi-season RAPM when generating the weights, as that increases predictive accuracy. As most of NBA.com’s advanced stats go back to 2016-17, we will be using RAPM from 2016-17 to 2025-26 for this exercise.

Since we are now using 10 years of RAPM, the individual data needs to match that same time frame: For each player in the sample, we will want to compute their 10-year “averages” across the same decade.

And because 10 years is rather long, I am using an age adjustment in RAPM, which boosts the ratings of players who play with really old or young teammates.

Take, for example, LeBron James and the Lakers. Because he is seen as a single entity across the 10-year dataset, we have to adjust his current teammates’ ratings to account for the fact that he’s 41 and not as good as he was for most of the past decade — the system gives his teammates a boost to avoid penalizing them for his decline.

As I showed in a previous article, an offensive player’s influence on the outcome of a possession is greater than that of a defender. Thus, using different penalization values for offense and defense makes sense.

Here’s what our setup would look like if we used 2016-17 to 2025-26 per-possession stats and total RAPM:

Weighting individual observations

No matter what time frame you select, your population of players will have a wide range of minutes played — from a lot to a few.

Obviously, Nikola Jokić should have more influence when generating the weights compared to, say, Anthony Bennett: Jokić’s RAPM is more accurate (with lower standard error), and we also have a more accurate grasp of his “true” career skill for each stat.

Simply from a machine learning standpoint, we want to rely more strongly on more accurate observations.

In the past, this situation was often dealt with by choosing a minutes cutoff, removing those players below the cutoff from the sample, and weighting the remaining observations equally.

But it’s rather arbitrary to give a player with, say, 1,001 minutes the same weight as another player with 20,000 minutes, while another player with 999 minutes might not have any influence whatsoever.

Fortunately, there is a better way: Keep all of the observations in the sample and weight them according to their RAPM standard error — that is, the level of uncertainty we have in their RAPM estimate.

As with box-score statistics, RAPM estimates come with their own measure of uncertainty. We get the standard errors through a technique developed by Justin Jacobs.

For the players in our RAPM sample, the standard errors look like this:

As expected, the uncertainty goes down with more possessions played. But it isn’t a linear relationship: We first get a steep decline with the initial possessions, but the curve tends to flatten out with more possessions.

The player with the lowest standard error should be given the largest weight. In our case that is Dennis Schröder. He has not played the most minutes in the sample, but because he’s a journeyman whose career includes the entire 10-year span, he’s played with hundreds of teammates, making us more confident in his RAPM estimate.

A player with close to zero minutes should be given a weight extremely close to 0. Later, the weights will have to be normalized so that the average weight is 1. Schröder ends up with a relative weight of around 3.

What’s omitted

Box Plus-Minus (BPM) is the best-known public version of SPM, as it’s published at Basketball Reference.

In contrast to BPM, several aspects have been kept simple in our version of SPM:

• There is no positional or role adjustment.

• There are no team adjustments.

• No team-related stats are required.

Stat normalization

There is one final pre-processing step: normalization of the different stats.

Instead of inputting, say, per-100 estimates into the regression, we are instead transforming each column to have an average of 0 and a standard deviation of 1.

For example, let’s take Victor Wembanyama: His block rate2 is eight standard deviations above the average posted by an average NBA player.

This technique is called “z-scoring” and generally helps machine learning algorithms to perform better. The resulting setup then looks a little different:

Subscribe now

Results

While this data setup can — like others we’ve previously presented — be put into many different machine learning algorithms, using regression has the benefit of telling us the relative importance of the different stats.

We get two immediate results from running the regression:

• A weight for each stat — this can be positive or negative.

• An SPM rating for each player (for the 10-year period we chose).

Because we run the analysis for both offense and defense separately, we get two sets of weights.

The analysis says that being above average at the following stats leads to higher offensive impact: minutes played; and points, offensive rebounds and assists per 100 possessions.

Low turnover numbers — especially dead-ball turnovers — and low block numbers are also positive indicators.

That last one might be surprising, but it’s an example of how advanced metrics can find and employ indirect signals.

Players with fewer offensive skills often take a different path to the NBA — through their defensive impact. That’s especially true for bigs.

The regression picks up this selection bias: Players with high block numbers might otherwise not even be in the NBA — if not for their rim protection — as they tend to lack passing and shooting skills.

On the defensive end, forcing misses near the basket, defensive-rebounding 2-pointers, blocking shots that go to the defense and drawing charges are the biggest positive indicators.

Defensive rebounds off of free throws are a negative indicator, probably because they signify a tendency to “stat-stuff.” And three categories that make opponent fast breaks more likely are also negatives on defense — blocked shots against, live-ball turnovers and offensive rebounds.

Using these weights, we can easily compute 10-year SPM ratings:

That said, in-sample comparison of SPM vs. RAPM gives us a good visual representation that defensive SPM metrics tend to struggle more when estimating impact.

The most likely reason is that the individual stats we have are less capable of capturing actual defensive impact relative to how well offensive stats capture offensive impact.

This is especially true without NBA advanced and play-by-play data: The only defensive stats in the box score are rebounds, blocks and steals.

Fluff vs. Fiber

These charts also help us identify high-fluff and high-fiber players:

Those far below the line tend to have a significantly higher SPM rating than RAPM. When the difference is extreme, that indicates the player might be doing things that can be counted but doesn’t “do the little things” that lead to winning.

Players above the line, on the other hand, tend to get underrated by SPM — their plus-minus often outpaces their countable contributions.

Single-season SPM

Using the weights, we can also compute SPM ratings for shorter time periods — single seasons, for example.

For the 2025-26 regular season, pure SPM ratings have these players at the top:

How to combine SPM with RAPM

Combining both methods gives us the best of both worlds: SPM is less susceptible to noise, and ideally RAPM captures what SPM has missed.

If we want to create a metric that’s optimized for predicting the immediate future, here is the approach that I found to be most accurate:

Employing the SPM impact estimates as Bayesian priors, we use about 2-3 years of lineup data to compute “prior-informed RAPM.”3

How much deviation we allow from these priors is ultimately an empirical question: If our SPM was perfect, and our lineup data was very noisy, we’d allow zero deviation (infinite penalization). With less accurate SPM and more accurate lineup data, the penalization gets reduced.

DIY

For those readers who want to run the analysis for themselves, we’ve compiled the data in a Google spreadsheet.

This includes 10 years of box-score, play-by-play and advanced data from NBA.com.4

Also included are the RAPM estimates and the corresponding standard errors.

I was very high on Collin Murray-Boyles, ranking him third on my final big board before the 2025 NBA draft.

His rookie season was a bit bumpy, however. An elbow injury and a recurring left thumb injury forced him to miss 25 games and presumably made shooting more difficult.

But in this Raptors-Cavaliers series — featuring multiple All-Stars — he looks like one of the best players on the court, despite being just 20 years young.

What’s the special sauce that makes this unique rookie so potent in the playoffs?

Read more

We released the results of my 2026 NBA draft model earlier this week, and you can find those here.

While the model has some of the usual suspects near the top, it also ranks several prospects significantly higher than you’ll find on most NBA draft boards.

Does this mean the model is overrating certain stats? Does it have blind spots for particular deficiencies?

Or are the NBA scouts and GMs still caught up in biases, such as a prospect’s ranking in high school?

Let’s check the tape for answers on six names that deserve a much closer look.

Read more

Let’s dive into the results — which prospects my NBA draft model likes and dislikes, and why.

What the NBA draft model does

Read more

We asked 12 leading NBA voices on Substack:

NBA playoffs: The stakes are highest for ...

Check out their answers and subscribe!

Subscribe now

|

While conceding that I’m probably not answering the question correctly and perhaps should be focusing on James Dolan announcing to the world in January that this a championship-or-bust spring for the Knicks — or fixating on Victor Wembanyama’s Spurs vs. Playoff Inexperience conundrum — I find myself thinking a lot about LeBron James’ future.

The Lakers realistically have little shot in their Round 1 series against Houston without the injured Luka Dončić and Austin Reaves, so it’s not like LeBron is expected to magically overcome their absences. Nor is there something to prove legacy-wise at this point. Yet there has never been more uncertainty about what happens next for King James than there is going into this series.

The leaguewide presumption is that the 41-year-old wants to play another season ... but where? Is there a pathway to stay with the Lakers when parting company would give the purple and gold much-needed flexibility to build a more well-rounded roster? Warriors, Cavaliers, Knicks, Nuggets ... do any of those destinations make legitimate sense when it is also widely presumed that LeBron does not want to leave Tinseltown? Or could this (gasp) actually be the last time we see him on the playoffs stage?

This much we know: We’ll finally starting inching closer to getting some answers when this Lakers season is over.

Subscribe to The Stein Line

|

Most to lose: Celtics. They don’t want just one ring in the Tatum-Brown-Queta era.

2nd most to lose: Rockets. If they can’t beat the disabled Lakers, why pay attention?

Most to gain: Nuggets. Nobody expects them to beat both OKC and the Spurs.

Something to eye: Detroit’s 3-point shooting. If it arrives, Pistons can win the East.

Subscribe to The Morning After

|

The stakes are highest for … Victor Wembanyama.

I don’t see a plausible path to a fifth ring for LeBron James or a first for an injured Luka Dončić. Meanwhile, a second ring for Shai Gilgeous-Alexander, Nikola Jokić, or Jayson Tatum wouldn’t boost them that much historically.

Wemby, however, has a rare chance to join legends Bill Russell, Magic Johnson, and Tim Duncan as an early-twenties NBA champion. If the thrilling Spurs win it all, the 22-year-old French phenom would claim his first Larry O’Brien Trophy well ahead of Michael Jordan (age 28), LeBron James (27), and Stephen Curry (27). What better way to launch a GOAT candidacy?

Subscribe to Ben Golliver

|

For the Cavs these playoffs feel existential. Perhaps a better framing: This group will decide whether it stays haunted or performs a group exorcism.

Some might say the Spurs and Wembanyama, but they get to exist in the bucolic realm of zero expectations and high skill. Then there’s the Pistons, the “prove it” team of this postseason.

Ultimately I’m picking OKC — and SGA. Repeating is always arduous, but this is a team that went from darling to target so fast, SGA included, that the eagerness to knock them out is palpable.

This was genuinely tough to narrow down; it’s good to have so many stakes!

Subscribe to BASKETBALL FEELINGS

|

The Eastern Conference was regarded as wide open thanks to injuries to Damian Lillard, Tyrese Haliburton, and Jayson Tatum. This created a ton of pressure for the remaining healthy teams in the East. We’re already bracing for potential changes in Orlando after falling below the lofty expectations the Magic raised by trading for Desmond Bane. The Knicks could also be looking to modify their roster if they can’t get back to the East Finals.

But perhaps no team has as much pressure on them as the Cavaliers. It feels like anything short of a Finals appearance could result in a major shakeup. It could be in either direction: Do they consider offering Evan Mobley for a superstar like Giannis Antetokounmpo? Or does Donovan Mitchell ask for a trade? There’s too much at stake after trading Darius Garland for 36-year-old James Harden.

Subscribe to Third Apron

Jake Fischer | The People’s Insider | Contributor to

I hate to pose as any grim reaper, but there are just too many good teams with high expectations to the point some club, maybe more, is inevitably going to find themselves at home after Round One and face some form of reckoning — whether that be changes to the roster, coach or front office.

The loser of Denver-Minneapolis, for example, very much entered this season with eyes on the title, let alone another conference finals trip. The same goes for the Lakers and Rockets. A lot of league personnel think Atlanta can topple New York.

What then?

Subscribe to The Stein Line

|

No matter how many angles I look at this from, I can’t see a bigger potential drama than an early exit (first or second round) for the Knicks.

On the positive side: Sure, if the Thunder win a second straight NBA title, it would elevate them because of their historic two-year dominance; if Wembanyama won with the Spurs, it would raise his global stature; for Boston, it would be a legendary season …

But any team or player has an excuse or a reason to survive not winning in 2026. Except for NY, obviously.

Subscribe to 82 partidos

|

The NBA desperately needs a fantastic playoffs for the Association’s own sake.

The Terry Rozier and Chauncey Billups gambling scandals. The still-lingering Kawhi/Ballmer/Aspiration stench. The endless complaints about tanking and the 65-game rule. An unprecedented barrage of injuries. Unceasing debate about whether the league is thriving or dying.

Despite an excellent on-court product, the NBA has been almost completely overshadowed by one publicity crisis after another. You know how best to remind people that this sport deserves love and attention? By playing some good friggin’ basketball.

The play-in has been a fantastic start. Here’s hoping for two more months like this.

Subscribe to Basketball Poetry

|

I’m excited for these playoffs for two reasons:

1. The postseason signals the end of the droning on about the problem de jour. In this case, I won’t have to hear about tanking (at least until the draft lottery) — thank you to whomever you constitute as God.

2. Wemby, SGA, and Jokić are looking to stake their claim as the game’s best, with the latter two inserting themselves into conversations that cast them amongst the all-time greats — even if “ring culture” has made championships an overblown metric for evaluating individual greatness.

Subscribe to Live. Breathe. Ball.

|

The New York Knicks.

They fired Tom Thibodeau, a great coach who changed their culture and took them to the conference finals, because that wasn’t good enough. The new coaching staff inherited a roster with Finals-or-bust expectations, set not by media or rivals, but by the organization itself.

Jalen Brunson is elite. The supporting cast is deep. The Garden is electric. But New York doesn’t do moral victories. If this team doesn’t reach the NBA Finals, the roster may get blown up. No pressure, though. It’s just New York.

Subscribe to NBA con Contexto

|

One player who qualifies is not in the playoffs at all: Trae Young.

The Hawks have absolutely surged since removing Trae from the equation behind the efforts of Jalen Johnson and a career-best version of Nickeil Alexander-Walker.

And yet, nowhere will the Trae trade be brought up more than in the playoffs, when the Hawks face up against Trae’s old prey, the Knicks. If this Atlanta team can make a surprise run, it puts a lot of reputational pressure on Trae, who has been addition by subtraction lately.

Subscribe to JSuttHoops

|

The Knicks’ entire starting unit will be on the wrong side of 30 before long. And their salary situation and lack of draft picks make it unlikely they’ll be able to add high-impact players without trading away core pieces.

Meanwhile, with Jayson Tatum not yet at full strength and the Pistons having limited playoff experience, this might be the best opportunity for New York to get out of the East.

Subscribe to 🏀 5x5

Like other advanced metrics, NBA draft models use machine learning methods to predict outcomes based on historical data. Most draft models help assign accurate weights to certain player actions. For example, a scout might know that an offensive rebound is more important than a defensive rebound, but how much more?

Draft models can easily internalize information on past prospects: Given that the entire point of the model is to make predictions about future prospects — while using past prospect data as a guide — a statistical model easily “remembers” which blend of player stats were a good predictor of success.

These models can also more easily adjust for age and strength of schedule, as we’ve shown here:

It’s no surprise then, that draft models can sometimes predict future stars — or busts — better than scouts and GMs.

But draft models aren’t perfect — human evaluators still retain advantages when it comes to evaluating fluidity, body language, injury proneness and other factors.

While it’s hard to estimate how much each NBA team is influenced by draft model input, recent selections indicate that some teams value it highly. For example, Reed Sheppard — a 6’3” guard who didn’t score much in the NCAA — probably doesn’t go No. 3 in 2024 without draft model input.

Today we’ll show what such a statistical NBA draft model looks like, where the pitfalls lie, and what model-design decisions can be made. We’ll also supply some of the data, so that potential readers can try things out for themselves.

What is a draft model trying to predict?

Key decision: Choosing a target variable

When building NBA player metrics, such as xRAPM or BPM, most, if not all, modern metrics have settled on trying to predict the same thing: a player’s influence on the team’s point differential when the player is on the court.

Doing it this way has several benefits:

• We already imply, in the design, that we’re looking for difference-makers, not those who simply accumulate counting stats.

• The scale of the results is intuitive to understand: A player with, say, a +4 rating would help outscore the opponent by 4 points — over an average player — if he were to play 100 possessions.

With draft models, this decision becomes a little more muddy. As a draft model creator, you’re faced with several design decisions. Choosing a sensible target variable is the first one:

Option 1: Predicting per-possession NBA impact

In this category, there are several metrics to choose from, including xRAPM and BPM.

My personal preference here is to use a metric that has a box-score and a plus-minus component, such as xRAPM. The box-score component increases stability and makes the metric less susceptible to noise. The plus-minus component allows us to identify under-the-radar difference-makers such as Franz Wagner or Derrick White.

Our NCAA dataset contains every season since 2009-10. We’ll thus want to compute a 2011-2026 NBA metric to serve as our “ground truth.”

Our custom NBA metric has the usual suspects at the top: Jokić, LeBron, Curry, and so on. But these players aren’t relevant to the model, as they didn’t play in the NCAA after 2009-10, if at all.

Among the best players relevant to the model are Joel Embiid, Shai Gilgeous-Alexander, Dereck Lively II, and Dylan Harper. Some of the worst-rated players are Nick Smith Jr. and Jordan Hawkins.

Our data setup would look something like this:1

Among other downsides, one issue with predicting an impact metric is that some highly touted NCAA players who failed — such as Jahlil Okafor, James Wiseman and Anthony Bennett, who all rated -4 or worse — rip quite a big hole into our projections. With the average outcome being relatively far in the red, the projections for new prospects remain fairly low.

The fact that some of these former high draft picks were this bad is also not really relevant to us: Any player below somewhere around -1 should not find himself in the regular rotation of a winning team. So, whether he’s ultimately rated as -1 or -4 is moot: From the perspective of a drafting team, neither player should be in its long-term plans, and drafting such a player would count as a failure.

Option 2: Predicting “Wins”

The above metrics don’t take into account playing time — durability is not considered since the metrics try to predict per-possession impact.

If we wanted to account for playing time, we could use metrics such as Win Shares or VORP, derived from Box Plus-Minus (both can be found at Basketball Reference).

Both of these estimate how many wins a player has added to the team’s season total:

When I worked for the Phoenix Suns and Dallas Mavericks, my colleagues in the analytics department sometimes opted for this approach, but I think there is a significant drawback: Players on bad teams can get too much credit.

The issue is that some teams give their young players more playing time — look at the Washington Wizards, for instance. In turn, that leads to those players getting more credit, not because they are better or more durable, but just because of the opportunities they were granted. In some cases, the issue is the team is below average at drafting and/or developing players.

Should the model really boost players picked by losing teams like, say, the Wizards or the Sacramento Kings? It seems like we’d be introducing a bias, and in the wrong direction.

Subscribe now

Option 3: Predicting probabilities

Alternatively, we can try to predict a probability — for example, predicting whether a player will rate as above “0” in the impact metrics, such as xRAPM.

Cutoffs can be chosen at will, leading to different models and coefficients.

For the purposes of this article, I decided to predict four potential outcomes for each prospect:

1. Player impact below 0: This group is made up of players who shouldn’t be in a winning team’s long-term plans.

2. 0 to 1.5: Typically a starter on a good team or maybe the best player on a bad team.

3. 1.5 to 3: Often one of the three best players on a contender, peaking as a top-20 NBA player.

4. Above 3: High-impact players, including superstars.

I like this approach for a couple of reasons:

• It better encapsulates the variance of outcomes. Player A might be more likely than Player B to be above average, but B might be more likely to become a star.

• It solves the problem of extremely low-impact players in Option 1.

Does this approach imply that our target variable is just 1s and 0s — indicating whether a player’s impact has been above certain threshold, or below?

Machine learning algorithms tend not to love that particular data setup. It’s also not quite correct once we consider confidence in our estimates as measured by standard errors.

Take, for example, Cason Wallace, rated by the NBA impact model at +2.7. If we’re building a draft model that tries to predict if a player will surpass +3, then the target variable corresponding to Wallace would be a flat “0” when not considering standard errors.

But the +2.7 represents only our “best guess.” The level of uncertainty in that estimate is expressed by the standard error — in Wallace’s case the SE is 1.5 — creating a situation best displayed in the following graph:

Incorporating these standard errors, we attain a value for a probability that Wallace is above +3: around 43%.2

One further consideration with this model and Option 1 is what to do with players who have seen very limited time in the NBA.

Per-possession metrics might not be harsh enough on certain players such as Usman Garuba, a 2021 first-round pick of the Houston Rockets, who rates as a ±0 in our custom impact metric.

Should a player who totaled just 1,228 minutes in three NBA seasons really serve as a positive example? While the answer is probably “no,” the solution isn’t so straightforward and requires some arbitrary cutoffs. Do we consider a player with less than 500 minutes per season a failed draft pick? 700?

What variables do we want to use?

Draft models try to make predictions based on historical data. But there are, of course, different types of data.

Do we include NCAA per-game data? Per possession? Totals?

My personal preference is to use estimates for each player’s “true skill.”

There are also intrinsic variables we can add. Among those are:

• age

• height

• whether the player was born outside the U.S.

• whether the player’s dad (or another family member) has played in the NBA

And we can add variables such as the quality of the team and its strength of schedule.

Note that some of these statistics have to be normalized — for example, “z-scored” — while others should exist as simple dummy variables.

Another big topic of debate is to whether to include past “pick number” information in the model.

One reason to do that: For players like Anthony Edwards — who had so-so college stats — scouts can identify attributes that a draft model isn’t able to see. For the current draft class, the player’s current draft board (or mock draft) rank would serve as a proxy.

Incorporating past draft pick information would make this a bit less of a “statistical” draft model, since we would be essentially including past scout/GM rankings, and we will rely on similar draft board information to make predictions.

On the plus side, this makes the model more accurate when predicting future success, since not everything is captured by the stats, and scout/GM evaluations add valuable information.

Not including draft pick information means the model stays “pure,” but the predictions will be a bit less accurate.

Which route you want to choose depends on your goals. Here’s a (paraphrased) discussion between me and the Suns’ director of analytics in 2013:

Results

Which variables are most important?

Given that we are running three different models — for each group of players, based on their probability of becoming a star, a bust, and so on — we are getting three different set of coefficients.

Here are the most important ones:

• Being an international prospect is a huge plus in regards to having an NBA impact rating above zero, but its importance diminishes with higher cutoffs.

• Free-throw percentage and 2-point percentage are the most important shooting-related factors, and they remain relatively stable in importance.

• Assists and height remain important with any cutoff.

• Steals and blocks go in opposite directions:

  • Steals are important for becoming a >0 player, but they lose importance with higher cutoffs.

  • Blocks, on the other hand, gain importance with higher cutoffs.

• Lastly, draft pick number is extremely important and gains in relative importance with higher cutoffs, indicating that the draft model has an easier time identifying solid players than superstars.

Here are, according to the model, the strongest NCAA seasons since 2010.

These represent “in-sample” results: We are asking what the model’s output would be, using hindsight, if presented with the following players today.

The following chart shows how the group predictions can vary wildly from player to player. In this example, Karl-Anthony Towns and Cooper Flagg were “predicted” to have a similar chance of becoming a superstar. But KAT had a significantly higher chance of becoming a bust.

Meanwhile, Franz Wagner had a lower “bust probability” than Towns, but the model saw less upside for Wagner and suggested he would most likely end up in the “Starter” group.

While we will present results for the 2026 draft class in a separate article, here is a Google doc containing past and current data that you can use to build a draft model on your own, and make projections for the 2026 draft class.

For this redraft, let’s assume the situation is similar to an actual draft: Bad teams come first, and they value upside potential, even if it means making the riskier pick at times.

What’s changed is we have more info now, based on each player’s first NBA season. We know who has passed the eye test. And we can look at their advanced metrics — especially for players who have dug themselves a hole that will be difficult to climb out of.

For example, if we had defensive concerns about a prospect before the draft, and his NBA metrics peg him as one of the worst defenders in the league now, a lot of our initial optimism goes out the window.

On the positive end of the spectrum, players whose upside was assumed to be hampered by their athleticism, but who are breaking through that glass ceiling, will get a boost.

Here’s how I see this year’s rookie class:

The new top 10

Read more

The San Antonio Spurs are the second-best team in the NBA — they’ve suffered just two losses in the last two months.

Yet, without Victor Wembanyama on the court, they are actually getting outscored by their opponents this season.

With his unique physical stature and playstyle, some proclaim Wemby is essentially breaking the NBA. That’s not crazy: He’s doing things no one has been able to do before.

Given that he’s just 22 years old, it would appear the rest of the NBA has a problem.

But let’s drill down to what’s really transformed Wembanyama into an MVP candidate and potentially the NBA’s best player — and perhaps the greatest defensive force basketball has ever seen.

Here’s the breakdown of where he is smashing NBA standards, based on advanced stats and video:

Read more

You already know college stats play a vital role when it comes to the NBA draft — a prospect’s numbers indicate his strengths, weaknesses, and future NBA production.

What’s more, statistical draft models — built on college player and team stats — outperform the actual draft order in predicting future NBA impact.

But here’s the challenge: To make those stats work for us, we need systematic ways to compare players and their stats.

Let’s say two players sport almost identical stats, but one of them is a year older: How much extra credit should we give to the younger one?

And what if one player has played against a much tougher slate of opponents? We have to figure out how much of a boost he should get when projecting his future.

To answer these questions, today we will:

• Look at prominent members of the current NBA draft class and how their ages and strength of schedule compares.

• Explain how to design a model constructed from box-score stats, adjusted for age and opponents..

The results are rather surprising!

We’ll also note which 2026 draft prospects are most affected by these adjustments.

Cameron Boozer vs. Caleb Wilson

An example of why we adjust college stats

While the top prospects In the 2026 draft are freshmen, there are some sizable age gaps between them.

Take, for instance, Caleb Wilson and Cameron Boozer: Wilson was born July 18, 2006, exactly one year before Boozer.

While Boozer’s stats are exceptional to begin with, they are even more impressive when we consider that he’s projected to be the youngest first-round pick in this draft — he’ll be among the very few in the one-and-done era to be drafted before his 19th birthday.

Not only that, but Boozer has played against stronger opponents, given that Wilson’s season ended on Feb. 10 while Boozer is still going in the NCAA tournament.

But while it seems rather obvious that Boozer would likely have even better numbers if he were a year older, little research has been done to determine how much better, until now.

In other words, it’s unclear how large the effects of age and schedule really are, and how much exactly we should be boosting or dampening player stats as we evaluate their likelihood of success in the NBA.

But using regression methods, we can determine the size of the effect of age and the quality of opponents on box-score stats.

Today, I’ll explain how to design a model incorporating these effects.

The technical details

The regression design

In our series of technical articles, our most recent piece covered how to predict NBA player stats.

Here is a quick recap:

First, we create an X-matrix where the number of columns corresponds to the number of players. If we’re estimating a player’s 3-point percentage, the number of rows in X is the total number of 3-point attempts of all players.

For each 3-point attempt (=1 row), the corresponding column of the shooting player is set to “1”.

The results vector Y — which has as many rows as X — is set to “1” for makes and “0” for misses. For a player who made three of his six 3-point attempts, part of our X and Y would look like this:

Plugging (X, y) into a penalized regression setup, we get a more accurate projection for a player’s 3p%, in contrast to taking his current 3p% at face value.

Adjusting for aging effects

This regression setup is easily extended to account for aging effects:

1. We create additional variables in our X-matrix, denoting the possible (rounded) ages of the players.

2. For every 3-pointer taken, we activate the age variable that corresponds to the age of the player at the time of each attempt.

Say that our player’s six 3-point attempts came in different games: Three makes in a game early in the season when the player was (rounded) 18 years old, and three misses later when the player was 19.

Then the new setup looks like this:

Adjusting for strength of schedule

This adjustment works in a similar way:

• We first calculate each opponent’s strength.

  • This can be win percentage, point differential, or something similar to Basketball Reference’s Simple Rating System (SRS), which adjusts for strength of opponent.

    • We “bucket” the various opponent strengths — for example, we can round to the nearest integer.

• We add a new set of dummy variables to X, one for each bucket.

• For each shot, we activate the dummy variable which denotes the strength of opponent in that game.

Suppose the opponent in the first game was rated +5, while the second opponent was rated +6.

With this setup, we can estimate aging curves and the effect of opponent quality, for each of the box-score stats.

A quick note: All the players used in this analysis were eventually drafted. This creates a biased sample, and the results would probably look slightly differently if we included all NCAA players. On the plus side, the results will probably be more accurate when predicting stats of future drafted players.

The surprising results

For points per possession, the effect of aging on college players looks like this:

Unsurprisingly, the graph points upward, denoting that NCAA players tend to score more points with each additional year of age.

The surprising part is this:

We might intuit that younger players make bigger leaps, year to year. Any parent or coach has seen how quickly young people can improve in any facet of life.

But for college players, we don’t see that.

Instead of the big improvements coming early in college and then the line looking flatter for older players, the improvement shown is remarkably linear, suggesting that the effect stays constant through at least age 23: A player is expected to score 3 points more (per 100) with each additional year of age.1 2

Here are the aging curves for assists, steals, blocks, rebounds and 3-point percentage:

Interestingly, blocks don’t see any increase with age — perhaps that is because this is the stat most dependent on raw athleticism, which might peak around age 18. Another explanation could be that players learn to contest shots by standing or jumping straight up, which can stymie opponents while creating fewer fouls — and fewer blocked shots.

For the other stats, the analysis suggests that with each year of aging, an NCAA player tends to gain:

• 3 more points (per 100 possessions)

• 0.5 assists

• 0.1 steals

• 0.4 rebounds

• 1% of increased 3p%

And here’s a chart denoting the improvement, in percentage, over the baseline (18 years old).

Points and assists rise almost 60% from age 18 to 23.

Rebounds, steals and 3p% rise by around 15% over the same timeframe, while block rate sees a small decline.

For the effects of opponent quality, the graphs looks like this:

As expected, all of these stats decline when playing against stronger opponents.

Playing a team like Houston or Duke — compared to an average opponent — will reduce stats by these amounts:

• Points by around 5 per 100 possessions

• Rebounds and assists by 2

• Blocks and steals by 0.5

• 3p% by 2.5%

Armed with these results, we can answer this basic question: What would everyone’s stats look like if they were the same age and played against the same quality of opponents?

Subscribe now

How this impacts Boozer and Wilson

To make these adjustments, we first have to calculate each player’s (possession-weighted) average age, and their individual strength of schedule.

For Boozer and Wilson, it looks like this:

• Boozer — Average age: 18.5. Strength of schedule: +8.0

• Wilson — Average age: 19.5: Strength of schedule: +5.8

Here are the projected stats for Boozer and Wilson if both players had been exactly 19 years old for every game and were playing average NCAA opponents each night.

As you can see, Boozer gets a decent-sized boost in his “Resulting Stats” line, while Wilson’s age- and schedule-related adjustments mostly cancel each other out.3

... and some other prospects in this draft class

If we use these adjustments for projected first-rounders of this current class, these are the players who get the biggest bump:

• Darryn Peterson has played the toughest schedule, with an average opponent SRS rating of +9.8. His points per 100 possessions would thus see a bump of around +2.

• On the other end of the spectrum, Bennett Stirtz’ opponents are rated at only +5.5, followed by Joshua Jefferson’s (+5.7) and Caleb Wilson’s (+5.8).4

• Cameron Boozer, Chris Cenac and AJ Dybantsa would see the biggest positive age adjustment.

• Yaxel Lendeborg — with a possession-weighted age of 23 — would see his 27.4 points/100 drop to around 15 using the above method.

Read more

The 2026 NBA draft is set up to be a very exciting, top-heavy one, with three prospects who would all go No. 1 in a weaker draft.

But these three players are extremely different — in body type, playstyle and stats.

The teams that win the No. 1 and No. 2 picks will have a tough, risky question on their hands: Who should they take?

Let’s answer that question here.

Strengths

Read more