Model card

Following the structure proposed by Mitchell et al., Model Cards for Model Reporting, FAT* 2019. Every number here is measured in this repository and reproducible from it. Where a figure is absent, the reason is stated rather than the figure omitted.

Model details

What it is. Regularized Adjusted Plus-Minus (RAPM) for NBA players, estimated by ridge regression over possession-level stint data, pooled into overlapping three-season windows.

Version. Tracks the package version. The shipped tables are regenerated only when the pipeline changes, never silently.

Owner. AlphaNerdFx. Apache-2.0 for the code, and the shipped tables are derived values computed in this repository rather than redistributed source data.

Estimator. sklearn.linear_model.Ridge over a sparse design matrix built in src/pippen/rapm/design.py. One column per player, +1 when on offence, -1 when on defence, 0 when off the floor. The penalty alpha is selected by cross-validation with folds cut on games rather than rows, so no possession from a game appears in both the training and validation side of a split.

What the model is not. It is not the fusion model. A separate measurement error model over box-score metrics exists in src/pippen/model/fusion.py, was tested, and did not improve on RAPM. It is not shipped as a rating.

Intended use

Primary use. Comparing player impact per 100 possessions within the covered seasons, for research, analysis and teaching.

Primary users. Basketball researchers, students, and anyone who wants an impact metric whose construction they can read end to end.

Out of scope. Contract valuation, betting, player evaluation for employment decisions, and any use where a wrong answer about one player carries consequence for that person. The metric has no per-player uncertainty published, which makes it unsuited to decisions about individuals at the margin.

Factors

Covered seasons. NBA 2016-17 through 2024-25, nine seasons, in NBA labelling. The lower bound is where data.nba.com stops serving play-by-play. The upper bound is where the feed stopped being populated.

Windows. Seven overlapping three-season windows, from 2016-18 to 2022-24. Adjacent windows share two of their three seasons, so consecutive values are not independent observations.

Population. 5,427 player-window rows. A player appears in a window if they recorded any possessions in it, with no minutes floor applied to the shipped table, so the low end includes players with under 20 possessions whose estimates are almost entirely prior.

Known population gap. Three of 5,427 rows carry no player name, where the ESPN to NBA identifier crosswalk failed for ids 2882 and 204058. All three sit under 20 possessions. The rows ship rather than being dropped, so the gap is visible rather than silently changing the row count.

Metrics

Reported. offensive, defensive and total, all in points per 100 possessions relative to league average. possessions is reported alongside as the per-player precision signal.

Not reported, deliberately. No standard error, no credible interval. The only candidate was a bootstrap standard error on the ridge coefficients, and measuring it showed the standard error rises with possessions, correlation +0.79. A penalised coefficient with little data behind it is pinned near zero and therefore varies little across resamples, so the tightest bars would have sat on the least-supported players. ADR 0007 records the decision, ADR 0003 the measurement that forced it.

Population reliability. Split-half reliability of RAPM is 0.796, measured rather than cited.

Training data

Source. Play-by-play from data.nba.com, parsed into possessions and stints with pbpstats (MIT). Box scores and schedules from the hoopR-nba-data Parquet release, CC BY 4.0.

Not used. EPM is behind a $250/year paid tier and its values are not redistributable. DARKO publishes a CSV with no stated licence. Neither enters this project; both are compared by rank correlation only, from manually viewed public leaderboards. Basketball-Reference values are used for local validation and never enter a published artifact.

Preprocessing. Every table passes a pandera schema and four validation families before use: completeness, consistency, accuracy and temporal integrity. Possession counts reconcile against official box score totals.

One correction worth naming. The possession coefficient, the share of free throw attempts that consume a possession, is measured per season rather than hard-coded at the conventional 0.44. Every one of 25 measured seasons came out below the convention, pooling at 0.4178.

Evaluation

Validation gate. Spearman rank correlation 0.914 against an independently constructed stint dataset, against a pre-registered threshold of 0.85. The threshold was declared before the measurement.

Calibration. The fusion model's intervals cover 90.9 percent at a nominal 90. That figure applies to the fusion model, which is not shipped as a rating, and it is a floor on the true uncertainty because pinning the anchor also pins its residual scale.

The claim under test, and its answer. The project asked whether fusing public metrics weighted by measured reliability predicts next-season team net rating better than RAPM alone. It does not, and the difference is not statistically distinguishable (p = 0.171). Ridge and a tuned LightGBM over every available metric land in the same place, so the ceiling belongs to these inputs rather than to the method of combining them.

Quantitative analyses

Reliability runs against validity. Across the 15 measured box-score metrics, split-half reliability correlates -0.564 with correlation against RAPM. The most repeatable metrics are the least related to winning. Three-point rate is the most reliable and among the least related to impact.

This is why nothing in this codebase turns reliability into a weight. It bounds how much of a metric a measurement-error model may attribute to the latent quantity, and is never a multiplier. ADR 0004.

A one-factor model of box-score metrics recovers position. Fit a single latent factor and it comes back as player size, because rebounds and blocks load together for tall players. That is why the fusion model pins an anchor rather than letting the factor float. ADR 0005.

Ethical considerations

The subjects are professional athletes whose on-court performance is already public. No personal data beyond name, league identifier and public box score lines is processed.

The risk this metric carries is misuse against individuals. A number with no published uncertainty, applied to a player near a roster or contract decision, invites false precision. The out-of-scope list above is the mitigation, and the absence of an interval is stated everywhere the number is served rather than buried.

Defensive impact is captured badly by box-score inputs, which is a documented property of the inputs rather than a flaw introduced here. Metrics that lean on box-score defence, DBPM and PER among them, are excluded entirely.

Caveats and recommendations

• Read possessions before reading the rating. A rating with 200 possessions behind it is mostly the prior.
• Do not read adjacent windows as year-on-year change. They overlap by two seasons.
• Do not weight by reliability. The measurement says it runs against validity.
• Do not use this for decisions about an individual's employment or pay.
• The fusion result is negative and published as such. If you are looking for the combined metric, it is not here, and the claim under test explains why.