Modeling API

Hash modeling-owned source files for reproducible analytical artifacts.

Build positive-exposure RAPM stints with conserved possession shares.

Validate curated regular-season inputs and publish RAPM stints atomically.

Read and validate one regular-season RAPM stint dataset.

Require exact RAPM part integrity and row-level invariants.

Build RAPM stints directly from selected persisted per-game tables.

Adapt cached pre-envelope per-game tables into valid RAPM source rows.

Build historical RAPM stints while excluding invalid game-level inputs.

Historical curated partitions can contain a small number of games produced before a lineup or possession reconstruction repair. RAPM requires exact point conservation, so an invalid game is excluded rather than allowing it to contaminate a season-wide design matrix. The caller persists the returned game IDs as part of its run provenance.

Construct comparable RAPM stints under one possession-allocation policy.

Quantify how each policy changes possession-to-lineup attribution.

A possession is changed when a policy's exposure vector over distinct ten-player lineups differs from the reference vector. Reassigned exposure is total-variation distance with an additional removed-possession bucket, so excluded exposure counts fully rather than as half a possession.

Create expanding folds without dividing any NBA game date.

Tune, test, and refit null, team, and canonical one-number RAPM models.

Build RAPM stints, run all baselines, and atomically persist artifacts.

Require every recorded model artifact to match its manifest.

Hash Bayesian RAPM source files for reproducible model runs.

Fit, evaluate, and summarize the conjugate counterpart to one ridge RAPM.

Fit exact Bayesian RAPM from one validated canonical ridge run.

Require every recorded Bayesian RAPM artifact to match its manifest.

Hash diagnostics-owned source files for reproducible reports.

Refit full-season RAPM over a lambda path and compare player ranks.

Compare fixed-lambda coefficients across expanding season windows.

Refit RAPM on complete-game bootstrap samples.

Measure how broadly each player is connected to teammates and lineups.

Compare raw on-court net rating with the adjusted RAPM coefficient.

Screen influential stints and verify top games with exact deletions.

Refit RAPM under alternative multi-lineup possession policies.

Combine review-oriented player diagnostics into one wide table.

Run every RAPM stability diagnostic and publish immutable reports.

Require every diagnostics artifact to match its manifest.

Assign transparent review bands to the exposure-eligible top ranking.

Generate a review page and charts from one immutable diagnostics run.

Render the complete one-season RAPM case study.

Join posterior and bootstrap uncertainty and return all eligible plus initial top N.

Generate a Bayesian-versus-ridge case study from immutable model runs.

Render the complete Bayesian-versus-ridge case study.

Aggregate played game boxscore rows into stable season-level features.

Combine same-season outcomes, box features, and static player context.

Create target-season rows containing only lagged performance features.

Build an atomic multi-season panel and leakage-safe lag-one transitions.

Validate exact panel artifacts, hashes, rows, and temporal columns.

Hash sources that define player-season and transition features.

Create expanding target-season folds and one untouched latest holdout.

Select and evaluate an aging model without using holdout outcomes.

Train from a validated player-season panel and publish an immutable run.

Validate hashes, rows, holdout identity, and label-free player priors.

Hash the implementation sources that define the aging experiment.

Return sorted entity IDs appearing on either side of a signed design.

Encode positive and negative entities in a SciPy CSR matrix.

Map unique entity IDs to stable contiguous sparse columns.

Hash neural-owned sources for reproducible datasets and runs.

Orient single-lineup possessions by offense and exclude ambiguous rows.

Validate curated segments and publish single-lineup possessions atomically.

Read and validate one regular-season neural possession dataset.

Require exact neural possession files, hashes, and row invariants.

Map stable NBA player IDs to embedding rows, reserving zero for unknown.

Build neural possessions, train additive RAPM, and persist an atomic run.

Select epochs, evaluate the final test, and refit the full season.

Require every recorded neural artifact to match its manifest.

Tune, evaluate, and refit the regular-season Deep Sets model.

Select optimization settings, evaluate fixed seeds, and refit all games.

Validate a Deep Sets run and require its architecture contract.

Return the CPU-conscious default search budget for Deep Sets.

Build leakage-safe stage predictions from the canonical RAPM baseline.

Fit one frozen RAPM state per chronological model stage.

Read and validate the stage-specific RAPM prediction rows.

Read fitted coefficients and stage parameters for base prediction.

Require exact files, hashes, roles, and chronological boundaries.

Hash source files that define the stage-aware base predictions.

Build frozen RAPM predictions, train the residual, and persist a run.

Select optimization settings, evaluate fixed seeds, and refit all games.

Require every Transformer artifact to match its run manifest.

Apply a stored RAPM state in offense orientation.

Apply a fitted Transformer and return total and residual predictions.

Hash Transformer-owned model, training, and evaluation sources.

Return the CPU-conscious first Transformer search budget.

Train, evaluate, and persist the categorical player-state model.

Fit the defaults-first CatBoost protocol on an in-memory dataset.

Encode each player as absent, offense, or defense plus home offense.

Return feature names in encoded column order.

Predict possessions and count player exposures absent from training.

Require every recorded CatBoost artifact to match its manifest.

Hash CatBoost-owned sources for reproducible runs.

Hash evaluation-owned sources for reproducible comparison reports.

Evaluate fitted regular-season models on regular holdout and playoffs.

Build common cohorts and score each stored point-prediction model.

Score multiple models on one identical possession cohort.

Bootstrap paired candidate-minus-reference RMSE differences by game.

Render the canonical equations and current model comparison tables.

Require every recorded evaluation artifact to match its manifest.

Validate the forward-prior RAPM artifact contract.

Return optional-weighted mean squared error.

Return optional-weighted root mean squared error.

Return optional-weighted mean absolute error.

Aggregate stint net ratings into game margins before computing RMSE.

Aggregate offense-oriented possessions into eligible home game margins.

Return out-of-sample skill relative to a baseline MSE.