BBall Index · Analysis

Forecasting Four Factor Impact: What Actually Predicts Next Season

If carrying over past performance isn’t the answer, what is?

Last writeup I cautioned against assuming last year’s impact metrics will translate to next season, particularly when context changes. Catch up on that article first if you haven’t already.

Our primary equation is still the same: Talent + Context = Impact.

If impact is the output we want to find, then predicting it means forecasting the inputs. That’s what I did for all ten of the granular four factor impact stats we discussed last go-around, using the three things we can reasonably predict before a game is played: what a player’s skills look like, the roles he’ll play, and who he’s about to share the floor with.

WHAT I BUILT

I used the same sample of guys as before: every player with 750+ minutes in back to back seasons, 2013-14 through 2025-26, resulting in 2,907 player pairs.

For each of the ten stats (offensive and defensive versions of Shot Quality Impact, eFG% Impact, Turnover Impact, FT Rate Impact, and Rebounding Impact), I tried to forecast next season’s value using five families of information:

  • Impact history: The player’s ten four factor stats, LA-ORAPM and LA-DRAPM, and a few other impact components. This is the data I cautioned against rolling forward (by itself) in my last article.
  • Talent: Offensive Talent, Defensive Talent, and every individual skill grade underneath them. To predict 2026-27 talent data, I utilized outputs from our Age Skill Growth Curves model you can find here.
  • Physical, age, & foul rates
  • New team context: Weighted average metric ratings for the collection of teammates each player is expected to play alongside. A lot went into this, and rightfully so given how much it factors into the projections:
    • Note: I’m using 2026-27 roster info as of 9/18/26.
    • Projected rotation MPG and total minutes were calculated for each player through a multi-step process:
      • Projected games missed data was pulled from our Injury Tracker tool.
      • General team strength was used to estimate garbage time frequency and reallocate minutes to the end of the bench.
      • I used the last 3 NBA seasons of info to model how projected injuries reconfigure rotations.
      • Role changes were calculated based on evolving player skills, reasonable lineup and roster-construction limitations (hi, Sixers), and coach biases/tendencies regarding role usage.
      • Prior year coach biases related to minute allocations were also calculated to result in team specific tweaks for known coaches that demonstrated verifiable favoritism towards certain demographics (roles, offense vs defensive skill, age, etc.).
    • Expected minutes alongside each teammate was then projected for each player.
  • Shifts from prior team context. Part of the team context analysis was also looking at shifts from the prior context. What the previous roster looked like, changes in offensive and defensive roles, and flags for team change, coach change and roster continuity.

HOW FORECASTABLE EACH STAT IS

Here we see correlations between what actually happened the next season and 3 kinds of projections: rolling forward the player’s own prior value, using only his talent profile, and the best full model.

Dot plot of ten four factor impact stats. For each, the out of sample correlation when rolling forward the prior value, when using talent only, and for the best full model. Offensive Shot Quality Impact is highest at 0.85. Defensive eFG Percent Impact is lowest at 0.31.

I have 3 main takeaways:

5 of 10stats where the talent profile alone beats rolling forward the player’s own prior value
10 of 10stats where the full model beats rolling forward the prior value
2% to 11%reduction in forecast error versus rolling last year forward, depending on the stat

Talent alone beats impact history on five of the ten before a single piece of context is added. Both eFG% Impacts, defensive turnovers, defensive rebounding, and (barely) offensive shot quality. As discussed in the last writeup, taking the player’s skill profile with no plus-minus data at all can predict the impact number better than the impact number predicts itself.

Running the impact data back never beat the full model. The gains range from 2% to 11% reduction in forecast error versus rolling last year forward.

Most stats are clustered, but there’s one that’s far off in each direction. Offensive Shot Quality Impact is the only thing here that behaves like a trait. It’s mostly playmaking, and playmaking travels. At the other end, Defensive eFG% Impact tops out at 0.31 no matter what I tried. With how unpredictable it is, I’d caution against any interpreting Defensive eFG% Impact as a defender trait. It’s more a description of what happened.

WHERE THE INFORMATION COMES FROM

This visual gives you a good look at which models lean on which stat families in their forecasting. The pattern is clean and matches up nicely with what you’d guess if you thought about what each stat is.

Stacked bars showing, for each four factor stat, the share of forecast information coming from the player's own impact history, his talent grades, the roster he is joining, tracking and physical data, the roster he is leaving, and role flags. Offensive Shot Quality Impact is 54 percent own history. Offensive Turnover Impact, Defensive eFG Percent Impact, Defensive FT Rate Impact and Defensive Rebounding Impact draw most from the roster he is joining.

Stats about a player’s own creation lean on his own history. Offensive Shot Quality Impact is 54% own history and 20% talent, with playmaking talent the single most important skill. Offensive FT rate is the same way.

Stats about sharing lean on the new roster. New team context is the single largest source of information for Offensive Turnover Impact, Defensive eFG% Impact, and Defensive Rebounding Impact, and it’s effectively tied with the player’s own foul rate for Defensive FT Rate Impact. Thinking about why, your Defensive Rebounding Impact depends on whether the guys next to you box out. Your Defensive FT Rate Impact depends on who’s fouling. Your Offensive Turnover Impact depends on who’s handling the ball when you’re not.

Defensive turnovers are the one place talent dominates. Passing Lane Defense is the single most important input, ahead of the player’s own prior value, with Pickpocket Rating third. Talent is the largest family at 30%. Steals travel!

You’ll notice roles aren’t carrying much weight themselves. Offensive role and defensive role account for 0% to 2% of the information once the talent and context are in. There is real role effect, but it’s carried by the skills and the roster instead of the label itself.

THE PAYOFF IS FOR THE GUYS WHO MOVED

Split the sample by whether the player changed teams and we see the real value in this exercise.

Dumbbell chart. For five four factor stats, rolling forward the prior value versus the full model, for players on the same team and players on a new team. On a new team the own value approach falls to between 0.32 and 0.40 and the full model recovers to between 0.44 and 0.54.

Rolling last year forward for a player on a new team is a coin flip on most of these. The model doesn’t make them as predictable as a guy who stayed put, but it recovers most of the gap, and it does it specifically by looking at the new roster.

Remember, Talent + Context = Impact.

ROLES CHANGE WHAT YOU CAN FORECAST

I expected accuracy to differ by role. You can see some of that with these examples.

Two dot plots. Offensive Shot Quality Impact is forecast at 0.80 for post scorers and 0.79 for shot creators but only 0.40 for stationary shooters. Defensive Rebounding Impact is forecast at 0.62 for mobile bigs and 0.56 for anchor bigs but 0.28 for point of attack defenders.

Offensive Shot Quality Impact is 0.80 for Post Scorers and 0.79 for Shot Creators, but for Stationary Shooters it’s just 0.40. And that makes sense, given the relative opportunity for each of those roles to generate offense for teammates both as passers and also in terms of creating advantages.

I wonder if the big man roles being higher than those couple shooter roles has to do with shot quality on putbacks and how the bigs will generate more of those chances. Off-Screen Shooters, who can pull 2-3 defenders into defending an off-ball screening action, above other shooter types is logical.

Defensively, Shot Quality, eFG% and Rebounding Impact are most forecastable for Anchor Bigs and Mobile Bigs. They’re in best position to deter and disrupt attacks to the rim (highest impact on eFG%/shot quality) and to rebound, so that makes sense. Defensive Rebounding Impact is 0.62 for Mobile Bigs versus 0.28 for POA defenders. Turnover Impact flips, with POA and Wing Stoppers most predictable at 0.56 to 0.57.

The biggest improvements from these full models compared to using prior year impact data come with the lower usage roles. Movement Shooters and Stationary Shooters offensively and Low Activity defenders are the guys whose single season impact numbers are the noisiest, so borrowing from talent and roster context helps them the most. Offensive eFG% Impact for Movement Shooters goes from 0.10 from prior year data (basically nothing) to 0.27 with the model. Still not great. But something instead of nothing.

WHAT IT LOOKS LIKE NEXT SEASON

Let’s quit talking about methodology and get to the fun part, which is what the data says about next season.

I ran the fitted models on the forecasted 2026-27 talent values referenced earlier from our talent projection tool, which already account for age and skill development. Teammate context comes from our 2026-27 minutes projection, which simulates each team’s season and tells us who actually shares the floor with whom, using those same projected teammate talents. Roles are projected 2026-27 roles, which lets a player’s role move with his evolving skills, his new roster’s needs and his coach’s tendencies (about a third of rotation players are projected to change offensive role, and a third to change defensive role; roughly half change at least one).

A few examples of what the forecasts say and, more importantly, why.

LaMelo to Minnesota

His Shot Quality Impact offensively holds at +2.52 (from +2.61), since his playmaking travels with him. His Offensive Rebounding Impact goes from +1.60 to +0.31, which is what happens when you move from a roster that needed a guard to crash the glass to one with Gobert on it.

Cooper Flagg in year two

Here’s one place the forecast goes up across the board. Offensive Shot Quality Impact went from +0.97 to +1.41, eFG% Impact from +0.43 to +1.19, and Offensive FT Rate Impact from -0.86 to +0.24. This is the projected talent doing the work, with his Playmaking Talent grade projected to improve from 1.17 to 1.63, and his context is stable. That’s a good recipe.

Bam Adebayo

Bam’s context changes without moving. Miami adds Giannis, and Bam’s sharing stats dropped accordingly: Offensive eFG% Impact from +2.26 to +0.60, Defensive Rebounding Impact from +2.15 to +0.87. Some of that is regression on noisy numbers, but the rebounding piece is a result of Miami adding a second elite rebounder.

THE 2026-27 FORECASTS

A couple example callouts. Players projected for 1,500+ minutes.

Offensive Shot Quality Impact
Player Team 2026-27 forecast 2025-26
Luka Doncic LAL +2.96 +3.81
Shai Gilgeous-Alexander OKC +2.83 +2.57
James Harden CLE +2.80 +2.26
Cade Cunningham DET +2.76 +2.57
Stephen Curry GSW +2.72 +2.92
Nikola Jokic DEN +2.53 +2.98
LaMelo Ball MIN +2.52 +2.61
Jalen Brunson NYK +2.36 +2.17
Defensive Rebounding Impact
Player Team 2026-27 forecast 2025-26
Donovan Clingan POR +1.29 +1.52
Nikola Jokic DEN +1.29 +1.74
Victor Wembanyama SAS +1.23 +1.59
Isaiah Hartenstein OKC +0.99 +1.10
Neemias Queta BOS +0.95 +1.09
Chet Holmgren OKC +0.93 +1.84
Karl-Anthony Towns NYK +0.91 +1.59
Bam Adebayo MIA +0.87 +2.15
Dumbbell chart of the biggest projected swings in Defensive Rebounding Impact. Isaiah Collier from minus 1.83 to minus 0.55, Peyton Watson from minus 1.44 to minus 0.23, Grayson Allen from minus 1.71 to minus 0.56, LeBron James from minus 0.61 to plus 0.24, Alperen Sengun from minus 0.38 to plus 0.39, Josh Hart from plus 1.37 to plus 0.42, Jayson Tatum from plus 1.65 to plus 0.54, Brandon Ingram from plus 1.62 to plus 0.35, and Bam Adebayo from plus 2.15 to plus 0.87.

HOW TO READ THESE NUMBERS

  • They regress large numbers on purpose. For the noisy stats (FT rate, defensive eFG%) the forecasts push closer to zero than outlier values. That is the model being honest about a 0.3 to 0.5 correlation. If a forecast looks boring next to a wild 2025-26 number, the wild number is the one to distrust.
  • Every forecast has an error bar. The workbook lists the out of sample error for each stat. Shot Quality Impact is more predictable and therefore has a smaller error bar. Defensive eFG% Impact is not, so you’ll see larger bars.
  • Talent and roles are next season’s, impact history is last season’s. The talent inputs are the 2026-27 projections and the roles are the projected 2026-27 roles, including partial mixes. Impact history is 2025-26 because there is no other kind.
  • Some players are forecast on talent and context alone. Rookies and guys who missed all of last season have no impact history. They get the reduced model and are flagged as such. Those forecasts are weaker on most stats, but they’re a real estimate instead of a blank. Rookie talent stats are estimates based on historical rookie results by role/position for draft slot groupings rather than reflections of college/international play. We’ll plan to improve this facet later.

THE POINT

Impact is what you get when a specific set of skills meets a specific set of circumstances. Last time I showed you the impact number doesn’t survive the trip to a new situation. This time we saw what does: skills, and an informed look at the new circumstances.

For the stats that are really about a player (shot quality, steals), his talent profile is most of the answer. For the stats that are really about a lineup (rebounding, fouls), the roster each player walks into is the biggest single input, bigger than his own history. Either way, the impact stat by itself is the worst of the available options for projecting itself, and it’s worst exactly when you need it most, for the guy who just changed teams.

NEXT UP

The analysis doesn’t end here. Next time we’ll talk through additional coaching impacts, then how the player level stats roll up to team performance expectations.