BBall Index · Analysis

Your Impact Stat Isn’t Coming Back Next Year

Why talent data and known context should drive your projections, not last season’s impact

Scroll for 2 seconds and you’ll see displayed on our homepage: Talent + Context = Impact.

Not Impact = Talent. Forgetting context from that equation is a common error, and one that leads to misguided analysis. Impact data doesn’t show how good you are. It shows how much you move the needle for your team, and that depends on the context around you.

Players are ingredients to a recipe. How much does garlic powder elevate a dish? It depends on what the other ingredients are! Yes, you’ll have your chocolates and your bacons where no heavy analysis is needed, but most NBA players are more rosemary than they are Reese’s.

When I refer to Talent, I mean the BBall Index metrics directly measuring performance in key skills (3PT Shooting, Playmaking, Rim Protection, etc.) that are treated in an attempt to remove situational impact factors in the pursuit of developing something situation-agnostic. Read about how we aggregate those offensively here. Defensively the ~same approach is taken, and those two stats are combined to form Overall Talent as referred to below.

“Impact” at a top line level refers to stats like LEBRON, EPM, DPM, BPM, etc. How much we’d expect a player’s presence to shift the team net rating.

But there are more granular impact stats. Ones you can find on our site and others. Stats like 4-factor impact, impact on transition frequency, and impact on opponent rim frequency (rim deterrence). They’re all useful to describe the past, but let this be your warning about trying to forecast the future with them.

Now that free agency is ~done, the go-to thing to do across fandoms in the league is to learn about your new players. Productivity, impact, etc. Some keep it to PRA. Others high dive into pools of granular data. Information on last year is good, but some is better to predict next year than others.

If you pull up a player and see they were a +2.5 guy last year in LEBRON/DPM/EPM/etc. and assume they’ll be a +2.5 guy next year, give or take a nudge for age, you’re going to get let down.

But look at the equation again. Impact is the output. It’s what you get when a specific set of skills meets a specific set of circumstances. Roll the output forward and you’re assuming the circumstances came with it.

And in general, they usually didn’t (I checked). For the shiny new object that just signed with your team, we know they didn’t.

WHAT ACTUALLY CARRIES OVER

I pulled 13 seasons of data, 2013-14 through 2025-26, and looked at every player who logged 750+ minutes two seasons in a row (2,895 guys). Then I asked the simplest possible question: how much does each impact metric look like itself one year later?

A metric measuring a skill perfectly that is completely agnostic to situation will have a higher correlation. It won’t be a perfect 1.0 relationship with no movement, because age and other fluctuations in performance are factors. And if it’s measuring noise and situation and not a trait we’d attribute to the player themselves, expect much lower numbers (which would be a big red flag if you want to be predicting the future).

Here’s the year-over-year correlation for each:

Year over year correlation for each metric on a number line. DPM 0.85, Overall Talent 0.81, BPM 0.78, LEBRON 0.77 and EPM 0.75 cluster at the high end while LA-RAPM sits at 0.50 and RAPM at 0.41.

DPM is by design a multi-year metric, so a higher value makes sense. Single year versions of LEBRON and EPM were chosen for this analysis, though ones using multi-year samples also exist.

The biggest takeaway is just how large the gap is between the RAPM stats and everything else. Raw RAPM is the least stable thing in the entire dataset. A player’s raw plus-minus impact estimate has a 0.41 relationship with his own number twelve months later. The luck adjusted version, Luck-Adjusted RAPM (LA-RAPM) buys you about nine points of that back (which is real and useful!), but still leaves you at 0.50.

Scatterplot of player RAPM this season against RAPM next season, a wide diffuse cloud with a correlation of 0.41.
Scatterplot of player Overall Talent this season against Overall Talent next season, a tight band along the diagonal with a correlation of 0.81.

Split this out by just offense or defense and you get the same story. Offensive Talent sits at 0.83. Defensive Talent, the single most stable defensive number of the whole sample, is at 0.87. Offensive RAPM is 0.38. Defensive RAPM is 0.35. HUGE gaps, and I’ll point out the values for offensive/defensive RAPM are lower than the overall. Get more granular (impact on offensive eFG%, defensive FT rate, etc.) and the range is 0.21 to 0.47. All quite poor.

EPM, LEBRON, and DPM each do utilize versions of RAPM for their calculations. BPM and Talent do not. All of these are much more stable than the RAPM stats. But the why is important.

Those metrics are stable because they’re anchored to box-score or tracking data priors that do directly track player production and move much less year to year.

TALENT PREDICTS YOUR IMPACT NUMBER BETTER THAN YOUR IMPACT DOES

Did that get your attention?

If you want to know a player’s LEBRON next season, the best single predictor in the dataset is not his LEBRON this season. It’s his Overall Talent! This ordering is true for both offense and defense.

A descriptive skill profile with no +/- data included in its calculations beats the impact metric at forecasting the impact metric. That should alarm you if your analysis process is solely “what was his impact last year.”

In terms of predicting team success when we aggregate values, RAPM ranks 1st of the metrics for the current season (descriptive) but above only LA-RAPM at predicting next season’s team success.

And if you’re seeking to forecast Offensive FT Rate Impact, Defensive Rebounding Impact, etc. granular 4-factor stats, we likewise see value in utilizing the talent data instead of rolling forward the current year impact stat to project next season’s rating.

CONTEXT SHIFTS

Here’s where it gets ugly and things separate even more. I flagged every player for four things:

  1. Did he change teams?
  2. Did he get a new head coach?
  3. Did his Offensive Role change?
  4. Did his Defensive Role change?

Then I compared the two extremes: guys going 4/4 vs guys going 0/4.

Predicting next season’s value in each stat, again using correlations:

Slope chart of each metric predicting its own value the next season, for players with no changes versus players with a new team, coach and roles. Overall Talent drops least at 18 percent while LA-RAPM drops 56 percent and RAPM 62 percent.

Everything degrades, but the RAPM stats plummet when context is disrupted.

Impact stats aren’t just noisy. They’re noisy in a specific way. They carry the fingerprints of a situation, and when that situation goes away, so does most of the number.

Kentavious Caldwell-Pope is a great example. In 2021-22 as a Wizard his O-RAPM was -1.18. The next year in Denver that spiked to +2.52. Same guy, same listed role, same skill profile (his Offensive Talent went from -0.28 to -0.44). Nearly four points of “impact” appeared out of thin air, and every bit of it was Denver’s scheme and roster and his fit within each.

If you were projecting KCP off his Wizards impact number, you missed. If you were projecting him off his shooting and defensive talent plus a hard look at who he was about to play next to, you didn’t.

SO JUST USE MORE YEARS OF IMPACT DATA, RIGHT?

That’s the standard fix, and it’s an understandable approach. A bigger sample size reduces noise. So let’s test it.

I built minutes-weighted versions of each stat so we could see 1/2/3/4/5-year versions of each for the 754 players where 5 years of prior seasons were available (and we could see the following year).

If we’re attempting to forecast RAPM it looks rough no matter what you use or how far back you go to grab more sample size.

Line chart of trailing window length against correlation with next season RAPM. Every metric lands between 0.43 and 0.49 and RAPM itself peaks at 0.475 only after four seasons.

Against a next-year LA-RAPM target, we get the same result: RAPM peaks at 0.50 and LA-RAPM at 0.55, and neither ever reaches the 0.575 you get from a single season of talent data. Against a next-year LEBRON target, one year of talent (0.82) beats five years of LEBRON (0.77).

Line chart of trailing window length against correlation with next season LA-RAPM. LEBRON and Overall Talent lead at roughly 0.58 with a one season window while LA-RAPM peaks at 0.548 only after four seasons.

Five seasons of plus-minus data still doesn’t buy you what one season of talent data does.

And now the part I didn’t know or expect until running these numbers: Multi-year talent barely helps at all. Two years is a hair better than one. Three is a wash. Four and five are worse.

Impact metrics need three to four years of averaging before they’re usable. Talent metrics peak at one to two. And by the time you’ve stacked enough seasons of impact data to trust it, you’ve averaged across so much change that you’ve blurred the player you were trying to project in the first place. That’s fine for a 30-year-old on his fourth year with the same team that’s at the only(?) part of the age growth curve where we don’t expect too much change over a 4-year stretch. It’s useless for a 22-year-old on his second team.

Diving a level deeper, each 4-factor impact stat for both sides of the court show similarly low values, other than Shot Quality impact offensively. Impact on opponent eFG% in particular is incredibly unstable year to year.

Line chart of the offensive four factor components against trailing window length. Offensive impact on teammate shot quality sits far above the rest at roughly 0.85 while the other four sit between 0.45 and 0.56.
Line chart of the defensive four factor components against trailing window length. No component clears 0.54 and defensive eFG percent impact sits far below everything at 0.28 even with a five season window.

Time-decayed metrics, which consider a player’s entire career and weigh towards recency, are an effective methodology you’ll find in our database for some stats and are a key element of DPM’s methodology. Had I included the multi-year versions of EPM and LEBRON above you’d see high stability as well. While this approach is useful in stabilizing, the reason that’s the case is only a ~third of each metric’s value is from the current season (at a 700-day half-life).

The exact breakdown depends on how many years of data a player has, but by comparing back to back seasons you’re going to be mostly (by %) comparing the same number against itself. It’s very stable, but it’s stable because the ingredients are mostly the same rather than because what it’s measuring is consistent. Even if you fully scramble which players have which RAPM numbers each season (so it’s pure junk) and then re-run time-decay RAPM at a 700-day half life you get 0.64. Some stability is there, but because it was re-using the same inputs rather than because there’s consistency.

And time-decay stats carry the old context forward, which leaves us in the same predicament when that context does change.

WHO THIS ACTUALLY HITS NEXT SEASON

Let’s shift focus back to today. We’re in offseason mode looking to project forward. How many of today’s players you’d be analyzing are in a bucket where changes in context make using 4-factor impact data or RAPM/LA-RAPM data an exercise in futility?

I took every player with 750+ minutes last season, matched them to current 2026-27 rosters, and checked the three context variables we already know today. Of 314 rotation players:

21.7%are on a new team
37.6%will play for a new head coach, including 20.4% of the guys who didn’t even move
22.3%are landing on a roster returning less of last year’s minutes than the historical league median

Stack those up:

49.4%have zero known changes
27.1%have one
16.2%have two
7.3%have all three

Roughly half the league’s rotation is walking into at least one changed context variable, and one in four is dealing with two or more. And that’s before role changes, which we won’t know until games are played. Historically, 56% of rotation players see their offensive or defensive role shift year to year, and it’s 62% among guys who changed teams.

League-wide roster continuity is 74.8% of last season’s minutes retained, which is actually on the stable side historically. The Lakers are at 42.5%. Dallas and Philly are both at 57%. If you’re projecting anyone in those rotations off last year’s impact numbers, you are projecting a team that no longer exists.

WHAT TO ACTUALLY DO

I’m not telling you to throw impact data out. But we can be smarter about how we use it.

For 7 of these 10 4-factor (and shot quality impact) metrics, along with RAPM and LA-RAPM, you can better predict the next year’s value using talent data rather than the current year impact metric.

But the best way to forecast impact, whether it’s RAPM, LA-RAPM, or any of the granular stats, we need to think back to the start: Talent + Context = Impact.

In this case, a blend of multi-year impact averages and single-year talent. Then take it a step further by adding in our BBall Index teammate talent data in pertinent areas.

Going an additional step, time-decayed talent metrics (keep an eye out for a launch here soon) take a similar approach of incorporating as much info as possible while helping measure skill development/decay over time rather than assuming it from team results.

Talent is the part of a player that travels. Context is the part you have to re-scout every single summer. Impact is what happens when you put them together, and it is an output, not an input.

So when you’re making your season predictions over the next few weeks and you catch yourself reaching for last year’s 4-factor data or RAPM/LA-RAPM: stop and ask whether this guy is one of the 50% whose context just moved. If he is, that number isn’t describing next year. It’s describing a team that’s already been dismantled.

Instead, look to their talent profile and new roster (in that order).

In future posts I’ll break down specifics, with methodology and forecasted values for players based on their projected next-year talent data (from this tool) and teammate context.