Skip to content

Context Function Audit

This report reviews the response functions already fitted by Forward Portable-Matchup Contextual RAPM. It does not train a candidate model or claim that a visible curve is a causal basketball relationship.

For season \(s\), feature \(k\), and relative feature difference \(z\), the audited total component is:

\[ r_{s,k}(z) = \frac{f_{s,k}(z)-f_{s,k}(-z)}{2}. \]

This is the model's original orientation-symmetrized total contextual response, before the season-specific reference field turns it into portable composition and matchup-residual terms. It is therefore the appropriate object for choosing a future function family.

Review Protocol

Inspect the fitted response only inside its observed central support. A function family is a candidate for a future refit only when its response is smooth and stable across seasons and its shape has a defensible interpretation. Engineered interactions face a higher bar: a visually irregular interaction is not retained merely because it appeared in an earlier feature set.

The next modeling decision is deliberately editorial and prospective:

  • retain a linear term when the atlas is stable and approximately straight;
  • consider a low-complexity spline only for stable nonlinear responses;
  • consider a monotone constrained response only with a defensible direction;
  • re-linearize or remove unstable interactions before paying for a full refit.

Current Audit

Source artifact: forward-portable-matchup-contextual-rapm-2025-26-20260808T165815Z-0d9d87f0. The audit evaluates 29 saved seasonal states without refitting. Faint curves are individual seasons, orange is the historical median, and green is 2025-26.

The x-axis is each season's 5th-to-95th percentile of possession-weighted independent reference-unit differences. This makes curve shapes comparable across eras without treating raw feature scales as stationary.

Main profile-feature response atlas

Composition-summary response atlas

Temporal Direction Versus Scatter

Each point below is a season's signed central response contrast: the 95th-percentile response minus the 5th-percentile response. Orange is a least-squares time trend. A low R² means the annual movement is primarily scatter rather than a sustained directional evolution.

Temporal central-response contrast atlas

Feature Contrast trend / decade Linear R² Season SD Observed contrast range
Bottom-two three-point makes -0.05 0.00 2.08 [-2.37, +6.46]
Credible-shooter count -1.36 0.11 3.52 [-9.64, +7.21]
Diminishing defensive rebounding -0.22 0.02 1.17 [-2.71, +2.74]
Diminishing offensive rebounding -0.36 0.03 1.64 [-2.54, +4.77]
Imputed-profile count -2.31 0.31 3.52 [-13.22, +4.53]
Rebounding-by-usage +0.37 0.02 2.13 [-2.90, +6.54]
Replacement-profile weight +2.19 0.49 2.68 [-3.80, +6.78]
Shooter-by-passing +0.49 0.02 2.93 [-5.11, +5.91]
Shooting-by-usage -0.77 0.15 1.69 [-3.77, +3.22]
Top-two assists -0.37 0.01 3.48 [-4.63, +12.34]
Usage concentration +0.00 0.00 2.14 [-3.41, +4.63]
Assists -0.99 0.07 3.22 [-6.95, +8.28]
Blocks -0.15 0.01 1.68 [-0.44, +6.04]
Defensive rebounds -0.36 0.05 1.37 [-3.43, +2.98]
Offensive rebounds +0.26 0.01 1.98 [-1.30, +5.63]
Steals +0.56 0.06 1.92 [-0.65, +7.10]
Three-point attempt volume -0.03 0.00 3.09 [-9.07, +5.04]
Three-point makes +1.05 0.08 3.20 [-3.29, +9.73]
Turnovers -0.50 0.04 2.03 [-5.49, +3.48]
Usage events -0.22 0.01 2.66 [-2.15, +9.04]

Stability Summary

Median range is the within-season 5th-to-95th response range. Median curvature is the largest departure from that season's endpoint chord. Median turns counts direction reversals across the central response interval. Cross-season SD is the median seasonal standard deviation at matched percentiles. These are screening diagnostics, not automatic feature-selection rules.

Feature Type Median range Median curvature Median turns Cross-season SD Latest support
Bottom-two three-point makes composition summary 1.22 0.11 0 0.34 [-2.32, 2.31]
Credible-shooter count composition summary 2.23 0.47 0 0.93 [-3.00, 3.00]
Diminishing defensive rebounding composition summary 0.77 0.17 0 0.18 [-0.70, 0.71]
Diminishing offensive rebounding composition summary 2.54 0.22 0 0.30 [-0.72, 0.74]
Imputed-profile count composition summary 2.74 0.80 0 0.90 [-2.00, 2.00]
Rebounding-by-usage composition summary 1.64 0.13 0 0.38 [-0.42, 0.42]
Replacement-profile weight composition summary 2.23 0.62 0 0.19 [-1.00, 1.00]
Shooter-by-passing composition summary 2.16 0.25 0 0.59 [-45.00, 45.51]
Shooting-by-usage composition summary 1.26 0.14 0 0.28 [-1.15, 1.16]
Top-two assists composition summary 2.72 0.31 0 0.66 [-7.14, 7.16]
Usage concentration composition summary 1.87 0.19 0 0.41 [-0.08, 0.08]
Assists profile rate 2.45 0.26 0 0.63 [-9.97, 9.97]
Blocks profile rate 3.30 0.29 0 0.35 [-2.88, 2.91]
Defensive rebounds profile rate 0.93 0.12 0 0.26 [-8.12, 8.14]
Offensive rebounds profile rate 1.47 0.19 0 0.36 [-5.04, 5.14]
Steals profile rate 3.03 0.27 0 0.36 [-2.34, 2.34]
Three-point attempt volume profile rate 2.72 0.23 0 0.59 [-14.62, 14.67]
Three-point makes profile rate 2.47 0.30 0 0.58 [-5.80, 5.84]
Turnovers profile rate 1.29 0.13 0 0.36 [-4.60, 4.57]
Usage events profile rate 4.50 0.36 0 0.50 [-26.73, 26.28]

Immutable generated artifact: artifacts/analysis/context_function_audit/2025-26/context-function-audit-2025-26-15Z-0d9d87f0.