IMPACT: A TWO-SIGNAL FRAMEWORK WITH A TRACKING OVERLAY

MOTIVATION & RELATED WORK

Evaluating individual NBA contribution from public data has a structural problem: every observable outcome (points, rebounds, team wins) is shared across five players on the floor. Basketball is a deeply cooperative sport, and attributing credit to a single player requires either (a) isolating each on-ball action and valuing it in isolation, or (b) regressing outcomes onto player-presence matrices and letting the regression untangle who causes what.

Box-score-based models (Win Shares, PER, PIE) take the first path. They are interpretable and responsive to current form, but reward volume, under-measure defense, and are blind to off-ball impact. Regression-based models (RAPM, EPM, LEBRON) take the second path. They capture everything that happens when a player is on the floor, but suffer from teammate collinearity, wide small-sample uncertainty, and a lag behind role changes.

These two families fail in opposite ways. Our approach is to treat them as two independent estimators of the same latent quantity, combine them by inverse variance, and let each player's own data determine their weight. The result — what we call Impact — is the number we display in the main Impact column on the leaderboard and player pages.

PIPELINE OVERVIEW

Every night, for each game in the NBA schedule, the full pipeline runs in four stages:

What follows documents each stage in full. Everything is published: equations, coefficients, derivations, sample sizes. No black boxes.

SIGNAL A — BOX-SCORE POSSESSION VALUE

Basketball is a zero-sum exchange of possessions. An NBA possession produces, on average, 1.12 points. Every event during a possession either moves the expected outcome up or down. Signal A values each event by its possession-outcome delta, attributes it to the player who caused it, and sums the result into a per-game number: total_impact.

Shot-difficulty adjustment

Before we value a made or missed shot, we need to know how hard it was. Our shot-difficulty model is empirical: every NBA shot is classified into one of 121 types defined by subtype (dunk, layup, pull-up, step-back, etc.), distance bucket, and shot value (2pt or 3pt). The expected FG% for each type is the observed success rate of all such shots in our training set. No latent variables, no modeling assumptions.

Event coefficients

Non-scoring events are valued by the empirical change in expected possession outcome they produce. Each coefficient below is measured directly from league-observed outcomes, not calibrated.

Scoring & shot-making

Creation credit

Baseline normalization

Raw event-value sums are not interpretable on their own — every player accumulates credit in proportion to minutes played. We subtract a per-game baseline so total_impact measures production above or below what an average rotation player would produce in the same minutes:

SIGNAL B — RAPM

Ridge-regularized Adjusted Plus-Minus regresses per-possession point differential onto indicator variables for every player on the floor. The coefficient for player i is the number of points per 100 possessions the team's net rating changes when i is on the court, controlling for every teammate and opponent in the same possession. It captures everything — off-ball spacing, defensive rotations, screen-setting, pace effects — that box-score attribution can't see.

λ selection. We cross-validate over the grid [100, 300, 1,000, 3,000, 10,000, 30,000, 100,000] and select the λ that minimizes held-out prediction MSE on the season's possession matrix. Typical best-λ values land in the 10,000–30,000 range given ~200,000 possessions and ~500 active players.

Output scale. RAPM coefficients are expressed in points per 100 possessions. In the 2025-26 season the per-player distribution has mean ≈ 0 and standard deviation ≈ 0.5, with extremes around ±2.

Team-demeaning

A naive RAPM coefficient entangles individual quality with team quality. All OKC players, for example, play exclusively with other above-average OKC players, and the regression will inflate every OKC player's coefficient by the team's shared excess. We subtract each team's mean RAPM before blending:

Team-demeaning strips roughly 0.3–0.5 points per 100 possessions of "team quality" from each coefficient in a typical year. This is the trade-off: we lose a bit of signal for stars on excellent teams (their real contribution partially masquerades as team quality), but we avoid systematically promoting every player on a 60-win squad.

THE BLEND

Given two independent noisy estimators of the same latent quantity, the minimum-variance linear combination is inverse-variance weighted: the estimator with lower variance gets more weight. This is the standard Bayesian result for combining measurements. Critically, every player gets their own weight, derived from their own data — there is no global box-vs-RAPM tuning parameter.

Per-player variances

We compute two variances for each player, both expressed in z-score units so they are commensurable:

Inverse-variance weights

In practice the weight distribution spans roughly [0.3, 0.8] on the box side, with median ≈ 0.60 — so the typical player's Impact is ~60% box-score and ~40% RAPM.

Per-game Impact

PBP CREDIT (COMPONENT LENS)

Alongside the main Impact column, the leaderboard displays six PBP Credit components: scoring, creation, shot-making, defense, hustle, turnovers. These are a parallel decomposition of the box-score signal — same shot-difficulty table, same event coefficients, but computed event-by-event from the raw play-by-play stream rather than aggregated from the box. Each component isolates one dimension of contribution.

PBP Credit is a diagnostic view, not the Impact number itself. The Impact column is the blended published_impact (§5). The PBP columns are shown because a single summary number doesn't answer questions like "is this player's Impact coming from scoring or creation?" or "which part of the box signal is driving their rating?" The decomposition makes those questions answerable at a glance.

A player's total PBP Credit (sum of the six components) closely tracks their total_impact from the box signal — the two are computed from the same underlying framework — but they are not identical because PBP Credit resolves events at the play level where box-score attribution cannot (e.g. live-ball vs dead-ball turnover classification, assist-to-layup vs assist-to-3 creation value). Implementation: compute_game_pbp_credit() in nightly_nba_summary.py.

NBA INSIDE-THE-GAME TRACKING

The leaderboard also displays four columns from the NBA's Inside the Game optical-tracking system, built on the league's partnership with Amazon Web Services. These metrics use 29-body-point skeletal tracking at 60 fps and are computed by the NBA itself, not by Read The Game. We fetch them from the NBA Stats API and show them alongside our Impact number as an independent third view.

These columns are especially valuable for dimensions where the box-score signal is systematically blind — primarily off-ball defense, which the NBA's skeletal tracking picks up but public play-by-play does not. A player who anchors great team defense without racking up steals or blocks will look weak in Signal A and often mediocre in RAPM (because team defense is shared across the on-court five), but can still register strong Leverage-defense numbers from the tracking feed. Showing these side-by-side lets a reader triangulate beyond any single lens.

READING THE IMPACT NUMBER

Impact is expressed as points per game above or below what a replacement-level rotation player would produce in the same role. The 2025-26 distribution across qualified players (≥20 games, ≥15 minutes per game) has mean +2.22, standard deviation 6.82, max +29.38, min −8.21. The positive mean reflects the qualification filter — by restricting to regular rotation players we exclude the bottom tail of garbage-time minutes.

VALIDATION

Any modeling framework must pass empirical tests. We validate Impact along four dimensions:

Internal consistency of the box signal. Signal A is zero-sum by construction (Equation 3.6): summing total_impact across any game's ten players yields exactly zero. This means the box signal cannot silently inflate or deflate — every overrated player is balanced by an underrated one. The blended published_impact does not preserve this exact property (RAPM is a season-level coefficient applied to every game, not a per-game attribution), but the component that feeds it remains internally balanced.

Signal agreement. The two signals agree where they should. Across qualified 2025-26 players, the rank correlation between season-average total_impact and team-demeaned RAPM is strong in the upper and lower tails — stars show up in both, and deep-bench players show up negatively in both. Disagreement concentrates in the middle of the distribution, exactly where inverse-variance weighting is most consequential.

Year-over-year stability. A metric that measures talent rather than noise should be strongly correlated with itself across seasons for stable-role players. We track this correlation and publish it as the model accumulates multi-season career data per player.

Live ↔ final reconciliation. During games we compute a box-score-only approximation (Live Impact) from the NBA's live feeds. After the nightly pipeline runs, the final published_impact is available. The correlation between a player's final in-game Live Impact and their post-game published Impact is a running validity check on the live approximation. As this reconciliation log accumulates over the season it will be published here.

Cross-metric correlation with EPM, LEBRON, and RAPM-based public models is a planned addition once we have sufficient licensed access to their season-aligned data.

KNOWN LIMITATIONS

We document limitations openly. These are the cases where Impact will systematically diverge from a player's true contribution:

What the blend cannot fix:

RAPM is season-level. A player who changes roles mid-season (trade, positional shift, injury-return minutes bump) will lag in RAPM until enough possessions accrue in the new role — for them Impact over-weights the stale Signal B. Team-demeaning strips real star-quality signal along with team-quality inflation, so genuine superstars on elite teams are slightly under-rated by the RAPM half of the blend. Rookies and mid-season additions have low-information RAPM estimates and effectively rely on Signal A until their possession count approaches the league median.

To partially offset the off-ball-defense gap, we display the NBA's optical-tracking metrics (§7) as a third independent lens. A player whose real defensive value shows up in Gravity or Leverage-defense but not in our box signal is flagged for the reader.

DATA SOURCES

All inputs come from the NBA's public APIs. No proprietary feeds, no paid data licenses:

• NBA Stats API · Play-by-play endpoints, box-score aggregations, and the Inside-the-Game tracking series (Gravity, Leverage, Shot Quality, FG% Above Expected).
• NBA CDN · Live scoreboard, live box scores, live play-by-play (used for the Live Impact surface during in-progress games).
• Basketball-Reference · Historical foul-drawn and defensive-matchup supplementary data.
• Training corpus · 1,030,404 shot records across four complete NBA regular seasons (2021-22 through 2025-26) plus three postseasons.
• Canonical derivations · Signal A: calculate_net_impact() in nightly_nba_summary.py. Signal B: scripts/compute_rapm.py. Blend: scripts/compute_published_impact.py.

CITATION & CONTACT

Citation

For front offices, researchers, media

Impact is a published model, not a black box. If you're building a research pipeline, writing a book, or running scouting analytics inside a team, we're happy to discuss methodology, share derivations, or produce custom extracts. Game-level data in parquet format is available on request.