Updated November 17, 2026. I reran this study with every season from 2000-01 through 2025-26 and added a section on teams. The tables now cover all 26 seasons, pad toward each season’s own league average (threes have changed too much for one average to fit every year), and let you plug in your own numbers. The original post was published August 6, 2020.
Stabilization Rates
One of the most common issues NBA fans and analysts grapple with in analyzing players is when does a certain level of new performance become ‘real’, e.g., “Marcus Smart is shooting 54% from three after 4 games: has he had a breakout, or is this just noise?” Enter the concept of ‘stabilization rates’, which is the idea that a player needs some number of shooting attempts or possessions before they ‘own’ a skill (alternatively you can use DARKO).
Stabilization rates are a concept which were originally popularized by Russell Carleton (AKA Pizza Cutter) writing about baseball, but have come to basketball as well from Darryl Blackport, Krishna Narsu, Nathan Walker, myself (and again), and others. In short, the idea is that at some point in a season, the skill aspect of a player’s performance starts to outweigh the inherent randomness in player performance, and stabilization rates are an attempt to figure out when that point is.
There are a lot of different ways to calculate stabilization rates – Carleton and Blackport used something called a “Kuder-Richardson 21 reliability score”, which involves grouping plate appearances/shooting attempts into buckets, and looking for the relationship in those buckets. Narsu simply looked for the point in the season at which a team’s current performance has more than 0.5 R^2 with their final statistics in a given metric. I prefer a predictive approach, via the ‘padding method’, which I first read about from Justin Kubatko, but is commonly employed by Tangotiger as well, and I presume has some long storied history I’m ignorant of.
The Padding Method
Under the ‘padding method’, at any given point in the season, you ‘pad’ a player’s performance with some sample of league-average performance, and that’s your projection of a player’s talent going forward the rest of the season. To take three point attempts as an example, Duncan Robinson opened the 2019-2020 season by shooting 17 of 34 from three (50%). There’s presumably some randomness in that performance, so we ‘pad’ both the numerator and the denominator there with some amount of league average performance (call this number X), to get a projection in the form of:
expected_fg3_pct = (17 + X * league_average_fg3_pct) / (34 + X)
If the padding number for threes is 260 (more on how to calculate this later), and the league average three-point-percentage is .358, then our projection becomes:
expected_fg3_pct = (17 + 260 * .358) / (34 + 260) = .374
So Robinson’s 34 attempts from three have raised our estimate of his three-point shooting talent from .358 (league average) to .374 (good, not great shooter). I think that roughly aligns with our intuitions as well – 34 attempts just isn’t very many, so we can’t go too wild, but it’s a nice sign all the same.
The beauty and elegance of the padding method is that the amount of padding remains constant throughout the season. You don’t need to use different coefficients depending on how much of the season has elapsed. At all times – you just add 260 attempts of league average performance to a player’s shooting to get their future projection. Early on, the 260 ‘padding’ attempts totally dominate a player’s projection. But eventually, as the sample size grows, the player’s actual performance has a bigger and bigger impact.
Applying that to Robinson, you get the chart below, showing the progression of his projected shooting, and his actual YTD performance. The cadence is the same, but the projection is naturally more conservative.
After his last game, Aug 14, 2020: 270 of 606, 44.6%. Projection 41.9%, 70% his own shooting.
Note even the more conservative projections eventually agree Robinson is an elite shooter.
Padding Values Are Stabilization Rates
So what does the padding approach have to do with stabilization rates? Well – they’re actually the same thing. Before games on January 8, 2020, Duncan Robinson had taken 259 threes, and made 115 of them. At that point in the season, his projection was:
expected_fg3_pct = (115 + 260 * .358) / (259 + 260) = .401
Meaning at that point, roughly half his projection was the 259 attempts he’d actually taken, and half was the 260 padding attempts. After that game, the 260 padding attempts made up less and less of his projection, and his projection consisted mostly of his own performance. I emphasize ‘mostly’ here to reinforce that there’s nothing magical about the ‘stabilization point’, or about 50%. We still add 260 league average attempts to Robinson’s projection even after this point – they just have less and less relative impact. (I’m not a huge fan of the term ‘stabilization’ for this reason, since it implies after X attempts a player’s performance is “stable”, but I’ve given up fighting on this point).
I’m a big fan of the ‘padding approach’ as a way to address the stabilization question for its simplicity and directness. It doesn’t involve somewhat opaque statistical methods such as KR21 rates, and it directly answers the question most people want to know when they’re discussing stabilization: how do we project this player going forward?
The Numbers for Players
With that somewhat lengthy preamble out of the way, here are the padding numbers/stabilization numbers for the main basic and advanced box-score metrics, sorted by most ‘stable’ to least, along with what I suggest ‘padding’ each metric’s performance with.
Projection for the rest of the season: 37.6% League average is 36.0%. The player's own numbers count for 11% of this. They reach half at 263 3-point attempts.
| League avg 2025-26 | ||
|---|---|---|
| 2.0games | 22.3 | |
| share of shots taken from three | 4.8shot attempts | 41.5% |
| share of teammates' made shots he assisted | 19teammates' made shots | 15.9% |
| 26free throws | 78.3% | |
| 38possessions | 7.4 | |
| free throws per shot attempt | 54shot attempts | .264 |
| 60possessions | 8.8 | |
| 60possessions | 5.4 | |
| 71possessions | 10.4 | |
| 76possessions | 19.9% | |
| 88possessions | 17.9 | |
| 92shot attempts | 47.1% | |
| share of his team's misses he rebounds | 94possessions | 4.6% |
| 103possessions | 2.3 | |
| per opponent 2-point attempt | 104opponent 2-point attempts | 1.9% |
| share of opponents' misses he rebounds | 105possessions | 13.4% |
| 107possessions | 6.5 | |
| 1282-point attempts | 55.0% | |
| 160possessions | 1.0 | |
| 194possessions | 4.7 | |
| 2633-point attempts | 36.0% | |
| 264possessions | 4.0 | |
| 400possessions | 2.8 | |
| 671possessions | 1.7 | |
| per opponent possession | 673possessions | 1.7% |
| per possession used | 716possessions | 13.8% |
| 998possessions | 0.0 |
These were calculated the simplest way I know how – trial and error via differential evolution, finding the best ‘padding’ number to add to year-to-date performance for each player, after every game of the season, for every season since 2001, to best predict rest-of-the-year performance (about 637K rows of data).
As you can see, minutes are extremely stable (duh), since they’re almost entirely a coach’s decision and not a question of randomness/noise. On the other side, plus/minus data is extremely noisy, and you should regress a player’s performance by about 1,000 league average possessions going forward.
Note that three point percentage is relatively noisy at 263 attempts before your performance is mostly ‘real’ rather than noise. But even 263 attempts is a much smaller number than Darryl Blackport had found in his 2014 piece, where he concluded you needed 750 attempts for ‘stabilization’. For context, only 5 players have ever taken more than 750 three point attempts in an entire season, so this is a pretty big difference.
Note also that a few of these have shifted over the last six seasons, judging by how well each padding number predicts seasons it wasn’t fit on. Field-goal and two-point percentage now settle much faster, after roughly 65 and 80 attempts instead of 105 and 150, and turnovers per 100 possessions settle a little faster. On the other side, fouls, assists and three-point attempts per 100 possessions take longer. Most other stats are about where they were.
The Numbers for Teams
The same method works for teams, i.e., take every team after every game since 2000-01, and pad toward that season’s league average.
Projection for the rest of the season: +5.0 League average is 0.0. The team's own numbers count for 50% of this. They reach half at 9.9 games.
| League avg 2025-26 | |||
|---|---|---|---|
| share of shots taken from three | 324shot attempts | 3.6 | 41.5% |
| 933possessions | 9.4 | 0.0 | |
| 9.9games | 9.9 | 0.0 | |
| possessions per game | 11games | 11 | 99.8 |
| 13games | 13 | 50.0% | |
| 707rebound chances | 16 | 26.0% | |
| 1,656opponent shot attempts | 18 | 41.8% | |
| 1,726opponent shot attempts | 19 | .262 | |
| 473free throws | 20 | 78.3% | |
| free throws per shot attempt | 1,894shot attempts | 21 | .264 |
| 2,092shot attempts | 23 | 54.6% | |
| points per 100 possessions | 2,780possessions | 28 | 115.8 |
| 1,257rebound chances | 29 | 74.0% | |
| points allowed per 100 possessions | 3,189possessions | 32 | 115.8 |
| 1,7052-point attempts | 33 | 55.0% | |
| 2,955opponent shot attempts | 33 | 54.3% | |
| 1,786opponent 2-point attempts | 34 | 55.0% | |
| 3,499possessions | 35 | 13.8% | |
| 3,528possessions | 35 | 13.8% | |
| 1,7113-point attempts | 46 | 36.0% | |
| 9,010opponent 3-point attempts | 241 | 35.5% | |
| 16,837opponent free throws | 717 | 78.3% |
A few things stand out here. Win percentage needs about 13 games of padding, and point margin about 10, meaning that ten games into a season a team’s margin is about half ‘real’. That fits with the early-season strength-of-schedule post, which found about 8 games for SRS. (That one adjusts for schedule and covers different seasons, so the two numbers measure slightly different things.)
Interestingly, net rating settles faster than either half of it. Offensive and defensive rating each need about 28 to 32 games of possessions, while net rating needs about 9. Some of a game’s randomness hits both ends at once (e.g., both teams shoot hot, or the whistle is tight), and cancels out in the difference.
On the other side, opponent three-point percentage is mostly luck. It needs about 9,000 attempts of padding, which is roughly three seasons’ worth, and opponent free-throw percentage is noisier still, at roughly nine seasons’ worth. Shot profile, by contrast, is set early: a team’s share of shots from three needs less than four games.
Note that free-throw rates, for and against, now take much longer to settle than they used to, since teams have become more alike in how often they get to the line.
Other Ways to Do This
Note that there are other ways to do this: you could attempt to predict the next game’s performance instead of the rest of the season’s performance. You could also use career-to-date performance to project rest-of-season performance, or career-to-date performance to project rest-of-career performance. There are pros and cons to all these approaches, although they mostly matter less than you may expect (which is one of the virtues of the padding approach in the first place). I’ll explore the others in upcoming posts however.
How the numbers are fit: for each stat, the one padding amount that minimizes squared error in predicting each player’s or team’s rest-of-season rate, weighted by the rest-of-season sample, padding toward that season’s league average. Possession-based rates such as usage and rebound rate are padded in possessions.