The Intuition Behind the Metrics

Follow how NBA impact models turn box scores and lineup results into estimates.

8 chapters · guided study · actual box score + simulated scenarios

Who played best?

Orlando beat Minnesota. Paolo scored more than Franz or Desmond, but each player’s line tells a different part of the game.

Actual NBA data · April 8, 2026

Final · Kia Center · 2025–26 season

MAGIC132:MINNESOTA120

Source: NBA.com official scorer’s report. Selected lines; retrieved 2026-09-14.

What the box score recorded
Magic playerPTSREBAST+/−
Paolo Banchero2086+12
Franz Wagner1702+20
Desmond Bane1846+22

Simulated example using Orlando players · one separate possession

What happened on the floor?

A made shot can also depend on a teammate who never touches the ball. Choose an action to see who gets a statistic and whose work stays outside the box score.

The recorded outcome

The action behind it

The box score records many important outcomes, but not every action that caused them.

Box Plus/Minus (BPM) asks how much player impact we can estimate from the information available in the box score and team results. It gives those numbers context; it still inherits their blind spots.

BPM · follow the data

Documented structure · actual box-score excerpt · no BPM rating calculated

Step through the functions. Watch the inputs become the next stage’s data.

Step 1 of 5

Collect the record

Paolo’s 20 points come from seven twos, one three and three free throws. A season supplies many more records.

Data in

Two-point makes
7 × 2
Three-point makes
1 × 3
Free throws made
3 × 1
Other recorded events
8 REB · 6 AST
collect_box_scores()

Data out

Paolo · points
20
Rebounds
8
Assists
6

Function names describe the operations; they are not provider source code. Methodology

How would a real box-score model work?

Model weights are estimated from historical data against a target related to impact. BPM uses a long-term adjusted-impact target to calibrate its formula, then uses box scores, role and team results. Its scale is points per 100 possessions above a league-average player; zero is average.

A possession is a team’s turn with the ball. The “per 100” scale puts different amounts of playing time on a common footing. We will connect it to the scoreboard next.

Box-score models can include role adjustments, interactions and nonlinear relationships. Correlated statistics mean changing one number is not a clean experiment in what caused a win. No BPM value is calculated or republished here.

Basketball-Reference’s BPM methodology

How much credit belongs to Paolo?

Start with a stretch the Magic win by eight with Paolo on the floor. He shares that result with four teammates. Change one substitution and watch the team’s result.

Simulated example using Orlando players

One simulated stint

MAGIC28:MINNESOTA20

Plus-minus tells us what happened while a player was on the court. It does not tell us how much of it he caused.

On/off compares the team with him to the team without him. The players replacing him and the opponents faced in those minutes still matter.

Show the math: scoring rates and on/off

Offensive rating = points scored ÷ possessions × 100. Defensive rating = points allowed ÷ possessions × 100. Lower defensive rating is better. Net rating = offensive rating − defensive rating. On/off = on-court net rating − off-court net rating, using each group’s own possession count.

These are authored team scores, not predictions of the named players. Teammate and opponent changes select different simulated totals. The two-point control adds or removes one made basket at a time. Ten-game totals are a separate simulated sample, not one stint multiplied thirty times. Both teams are assigned the same possession count within each sample.

No individual ability values are known or printed. The names come from the dated roster snapshot in the methodology drawer.

Who should receive the credit?

Paolo and Franz start together. At first, every good result belongs to both. Step through the substitutions to find minutes that tell us something different about each player.

Simulated example using Orlando players · not a replay of April 8

Q1 · 12:00 → 9:36

MAGIC8:MINNESOTA4

When Paolo and Franz played

One lineup result tells us how five players performed together. Many overlapping lineups give the model opportunities to separate them. Different opponents provide another piece of that puzzle.

This is the idea behind adjusted plus-minus: compare the results while accounting for who shared the floor. Six stints help explain the question; they are far too little evidence to settle player value.

Adjusted plus-minus · follow the data

Calculated teaching example · simulated Orlando lineups · points per 100 possessions

Step through the functions. Watch the inputs become the next stage’s data.

Step 1 of 5

Read the scoreboard

In the first simulated stint, Orlando scores eight and allows four. The observed margin belongs to the whole group.

Data in

Orlando
8
Minnesota
4
Possessions
5
subtract_scores()

Data out

Scoring margin
+4
Possessions
5

Function names describe the operations; they are not provider source code. Methodology

Show the math: the same stints as a matrix

Each row is a stint. Each named player gets a column. +1 means playing for Orlando, −1 means playing for Minnesota, and 0 means sitting. The outcome is scoring margin per 100 possessions. A regression fits player contributions whose sums approximate those margins.

Player presence in the stints shown above

Filled: Orlando · outlined: Minnesota · blank: sitting

Simulated stint records
StintOrlandoMinnesotaPoss.Margin

Fitted player numbers are called coefficients. Perfectly shared minutes leave some of them inseparable. Because the design contains five players on each side, adding the same value to everyone changes no predicted margin; a reference constraint is required. Accounting for recorded lineups also does not prove causation.

Adjusted plus-minus background

How much should one lineup tell us?

Paolo and Franz keep sharing the same good results. Before deciding who deserves the credit, follow how the model uses those minutes and where it needs to be cautious.

Simulated example using Orlando players · simplified model

Now change the evidence

Keep the lineup results fixed and add more caution, or change the evidence the model gets. The points below are simulated estimates, not ratings of the real players.

Orlando estimates from the simulated lineup results
Simulated player estimatesDots show estimates. Whiskers show changes under specified disturbances to scoring, not statistical confidence.

Dot: estimate · whisker: response to small scoring changes
0 = average in this simulation; units are points per 100 possessions

Paolo and Franz as the model becomes more cautious

Paolo: solid · Franz: dashed · vertical: current setting

The second option adds invented starting expectations instead of pulling everyone toward zero. It does not add new game evidence.

RAPM · follow the data

Calculated teaching example · simulated Orlando players · fixed 48-observation dataset

Step through the functions. Watch the inputs become the next stage’s data.

Step 1 of 5

Collect lineup evidence

Retain who played, the scoring margin and the possession count. These are simulated observations, not real player results.

Data in

Named lineups
Orlando + Minnesota
Scoring outcomes
48 observations
Sample
2,400 possessions
prepare_lineup_rows()

Data out

Player-presence matrix
48 × 16
Margin rates
Points per 100
Observation weights
Possessions / 100

Function names describe the operations; they are not provider source code. Methodology

Show the math: ridge regression, priors and stability

Regularized adjusted plus-minus (RAPM) uses a penalty to resist extreme estimates. The penalty is called regularization; the particular squared penalty here is ridge regression. A prior is the starting value each player is pulled toward.

We minimize Σ (n / 100) × (y − Xβ)² + λ × Σ (β − prior)². X marks the ten players, y is scoring margin per 100, and n is possessions. The browser solves the resulting system. The caution slider changes λ logarithmically. Its presets are teaching choices, not validated NBA settings.

The dataset contains 48 simulated lineup combinations and 16 named players, eight per team. The chart shows Orlando; the table includes Minnesota. The evidence for each combination represents an aggregate when the sample control increases. Hidden synthetic effects generate the margins; those parameters are not estimates of these players’ abilities. Adding scoring margin with Paolo changes the observed targets in rows where he plays.

Default priors are zero. The optional informed priors are Paolo +3, Franz +1, Terrence Shannon Jr. −4 and everyone else 0. These are invented assumptions with a zero mean. If two players always share the court, the prior can determine a division that the data cannot establish.

Whiskers span the base fit and twelve refits. Refit k adds 12 × sin((row + 1) × (k + 1) × 1.71) × √(50 / n) to each target. This is sensitivity to a specified scoring disturbance, not a confidence interval. Shrinkage can produce a stable number without resolving missing information.

Production RAPM requires real possession data, additional context, an explicit sample window and tuning on held-out games. This model estimates total contribution only, relative to its own player pool. It does not establish causal NBA ability or roster fit.

Simulated coefficient and sensitivity range / 100
PlayerEstimateSensitivity range

Sill’s regularization paper

Which information enters the model?

A prior is a starting estimate. A hybrid model can use recorded production to set that starting point, then update it with lineup evidence.

Publicly described inputs and model family
  • Box score
  • Lineup outcomes
  • Team / opponent context
  • Tracking
  • Role / position
  • Prior seasons
  • Aging
  • Stat-specific reliability

Points / 100 possessions
0 = league average

Solid marker: documented input · dash: not established here as a direct input. Connections do not represent weights.

EPM · follow the data

Documented architecture · symbolic intermediate values · not provider code or ratings

Step through the functions. Watch the inputs become the next stage’s data.

Step 1 of 5

Collect player history

Box scores and tracking supply a history of player statistics.

Data in

Box-score records
Past games
Tracking records
Additional observations
collect_stat_history()

Data out

Stat histories
Shooting / passing / more

Function names describe the operations; they are not provider source code. Methodology

Compare methods across all dimensions
Model definitions reviewed 14 September 2026
DimensionBPMRAPMEPMLEBRONDPM
Core questionWhat contribution does the box score suggest?How does margin vary with the players on court?What impact is expected from this player now?What impact did a player have in this sample?What is the player expected to contribute next?
Model familyBox-score regression + team adjustmentPenalized lineup regressionSkill-based prior + decayed RAPMStabilized box prior + luck-adjusted RAPMDynamic projection system
OrientationRetrospectiveExplanatory; can be tuned for predictionPredictive (current main model)Retrospective seasonal impactPredictive
Time horizonSelected seasonChosen sample, often multiple seasonsAs-of date; career evidence decaysSeason (version described in introduction)Daily; core calibration is next-day
PriorsNo player-history prior in the seasonal calculationZero or an explicit informed prior; implementation-specificStatistical plus-minus from estimated skillsBox-score prior stabilized with offensive rolesEarlier projections updated with new evidence
Lineup adjustmentTeam adjustment; no stint regressionTeammates and opponentsTeammates and opponentsTeammates and opponents; luck adjustmentsOn/off added to Box DPM; context adjustments
Offense / defenseOffense; defense = total − offenseAvailable in offensive/defensive implementationsOffensive and defensive EPMO-LEBRON and D-LEBRONOffensive and defensive DPM
Typical stabilityLess lineup noise; inherits box-score blind spotsMore stable than APM; still sample-dependentHistory and skill stabilization reduce noiseRole stabilization helps small samplesUpdates depend on player and statistic
Major limitationDefensive positioning is poorly observed.Persistent lineup overlap limits identification.Model assumptions and incomplete observed inputs remain.Luck and role adjustments involve modeling choices.Future health, minutes and role remain uncertain.
TransparencyPublished formulaPublic method; implementation variesMethod described; full production model unavailableComponents described; not a complete reproductionArchitecture described; production details incomplete
Direct box-score dependenceYesNo, unless used as a priorYesYesYes

Descriptive: reports what happened. Explanatory: allocates observed outcomes conditional on a model. Predictive: estimates future performance. One method can serve more than one purpose; training targets, validation and time windows determine the use.

Paolo just scored 42. What changes?

How much should one game change what we expect next time? Follow the update, then try a longer run of games.

Simulated example using Orlando players · simplified model

Simulated game 1 · final

MAGIC149:MINNESOTA124

A conceptual illustration of sequential model updating, not the proprietary DARKO calculation.

Expected scoring after each simulated game

Dots: scored in each game, adjusted to 36 minutes · line: next-game expectation at 36 minutes

DARKO updates its expectations as new games arrive. It projects box-score statistics as well as player impact. Tonight’s box score describes a game; the updated projection describes what the model expects going forward.

DARKO / DPM · follow the data

Calculated scoring-update illustration → symbolic DARKO architecture · no DPM rating calculated

Step through the functions. Watch the inputs become the next stage’s data.

Step 1 of 5

Carry yesterday forward

Start with the teaching model’s existing scoring expectation. Actual DARKO maintains projections for many statistics.

Data in

Previous scoring expectation
24.0 points / 36 min
read_previous_expectation()

Data out

Expected tonight · 36 min
24.0 points
Other statistics
Separate projections in DARKO

Function names describe the operations; they are not provider source code. Methodology

Show the math: expectation, surprise and updating

The prior is the expectation before the new game. The observation is the new scoring line. Here we compare points at the same playing time: observation = points × 36 / minutes. The expected observation is the previous model mean. The residual, or surprise, is observation minus that expectation.

This small normal-updating example begins with variance 4, adds 0.08 before each game, and uses measurement variance 64 × 36 / minutes. Gain = prior variance / (prior variance + measurement variance). Updated mean = old mean + gain × surprise. Updated variance = (1 − gain) × prior variance. These are chosen teaching assumptions, not DARKO parameters.

More minutes can raise a scoring total without raising the scoring rate. During a 30-day absence, this toy model widens its variance by 30 × 0.08, with no automatic reduction in the mean. It then updates from the simulated return games. This does not model injury severity, availability or recovery.

No uncertainty interval from this toy calculation is presented as NBA forecast accuracy. DARKO’s full production system includes many statistics and different update rates; this scoring-only example does not calculate Daily Plus Minus (DPM).

Official DARKO methodology

Why metrics disagree

Choose a fictional player profile. Each point represents a different model’s hypothetical estimate, on the same points-per-100 scale.

Illustrative estimates across five model lenses

Same units, different targets. 0 = league-average reference.
Positive defense means points prevented, improving the scoring margin.

Invented values · 2026-09-14 · no real-player ratings

Exact illustrative values and comparison boundaries

Total = offense + defense here. Offense estimates points added; defense estimates points prevented. Providers may round separately. These values are authored examples, not outputs of the teaching ridge model or any provider. They demonstrate possible disagreements, not evidence that a metric favors a player type. Raw plus-minus and on/off are excluded because their reference points differ.

Which metric should I use?

Start with

These metrics are models answering overlapping but different questions with different evidence. Define the question, the time period and the comparison before choosing a number.

Check the offensive and defensive components, sample size, role and minutes. Use film and other evidence to investigate disagreement. Averaging several correlated metrics does not guarantee a better answer.

Methodology, sources and data

Actual NBA data: Chapter 1 contains three selected box-score lines from Orlando’s 132–120 win over Minnesota on 8 April 2026 (2025–26 season).Source: NBA.com official scorer’s report. Retrieved 14 September 2026; fields reordered and scoring arithmetic checked. No possessions or additional player results were reconstructed from this report.

Simulated NBA scenario: All other game results, floor actions and lineups are original teaching examples using players from that dated game roster. They do not describe those players’ actual performances. The selected Orlando names were also checked against the NBA.com Magic roster on 14 September 2026. Opponent names are from the April game snapshot.

Simplified model: The lineup calculations, regularized estimates and sequential scoring projection run in your browser under the disclosed assumptions. They are not published NBA player ratings. The comparison profiles are fictional.

Published metric snapshot: None is included. No provider ratings, proprietary weights or live data are loaded. Methodology descriptions were reviewed on 14 September 2026; the fictional comparison dataset is version 1.

  1. Daniel Myers · BPM 2.0 · reviewed 2026-09-14
  2. Dunks & Threes · Estimated Plus-Minus · reviewed 2026-09-14
  3. Kostya Medvedovsky · DARKO methodology · reviewed 2026-09-14
  4. BBall Index · LEBRON introduction · reviewed 2026-09-14
  5. Steve Ilardi · Adjusted plus-minus · reviewed 2026-09-14
  6. Joseph Sill · Improved NBA adjusted +/- (2010) · reviewed 2026-09-14

The small factual game excerpt is attributed to NBA.com; no logos, photographs or play-by-play dataset are reproduced. See NBA terms of use. Simulation provenance and calculations are described in each optional math layer.

Real ratings would need source-specific redistribution permission and matching dates, variants and baselines. Dunks & Threes’ terms require permission for public API-data display. No source’s public availability is treated as permission to republish its data.