Portable Context Attribution Audit

Value-Conditioned Aging HPM is intentionally predictive: it can place a lineup-level pattern in the portable composition term (h(U)) or the specific matchup term (q(U,V)), rather than forcing all of it into five player coefficients. This audit asks a separate accounting question: conditional on a specific unit, how does the frozen context function change as each player's preseason profile enters the unit?

The results are model attribution, not causal credit. They explain how HPM uses player profiles in its context layer; they do not prove how much a player caused the observed result.

For a whole-season player reallocation of frozen context, see Context-Reattributed RAPM. This Shapley audit instead explains a selected unit's composition and matchup ledger.

Method

The frozen 2024-25 context state evaluates a 2025-26 matchup as

\[ C(A,B) = h(A) - h(B) + q(A,B). \]

For composition accounting, each player in a unit is replaced by one of five distinct synthetic league-average profiles. Those profiles are possession weighted from the completed 2024-25 player pool. For every subset (S) of the five real players, (v(S)) is the portable score after players in (S) enter the otherwise-reference unit. Player (i)'s exact Shapley value is

\[ \phi_i = \sum_{S\subseteq N\setminus\{i\}} \frac{|S|!(|N|-|S|-1)!}{|N|!} \left[v(S\cup\{i\})-v(S)\right]. \]

Thus the five values add exactly to (h(A)-h(R)), where (R) is that synthetic reference unit. They do not necessarily add to (h(A)), because HPM defines (h(A)) against its empirical reference-unit distribution rather than against (R). The reference cancels when sides are compared:

\[ \sum_{i\in A}\phi_i^{\mathrm{comp}} - \sum_{j\in B}\phi_j^{\mathrm{comp}} = h(A)-h(B). \]

The matchup residual uses the same calculation across all ten players, with both sides initially set to (R). Its ten Shapley values add exactly to (q(A,B)). Therefore every player row has a signed contribution to the focal unit's edge and all ten rows reconcile to (C(A,B)).

HPM player rating + context contribution is included as a useful unit ledger, not a newly calibrated or leaderboard-eligible player metric.

Selected Units

These are frozen 2025-26 evaluations using realized units. The possession and game counts only describe how often the exact matchup occurred; they are not used in the frozen context score itself.

Unit matchup	Realized possessions	Games	Player edge	Composition edge	Matchup edge	Total context	Predicted net rating
Warriors unit at Spurs unit	37.3	2	-11.18	+1.40	-2.39	-0.99	-12.17
Rockets unit vs Clippers unit	46.5	2	+2.08	-1.21	+0.61	-0.60	+1.48
Knicks unit vs Hawks unit	36.5	1	+0.45	+0.47	-0.20	+0.27	+0.72

Warriors Unit At Spurs Unit

Focal unit: Stephen Curry, Jimmy Butler III, Draymond Green, Moses Moody, and Will Richard. Opponent: Harrison Barnes, De'Aaron Fox, Devin Vassell, Victor Wembanyama, and Stephon Castle.

Player	Side	HPM player rating	Player edge	Composition contribution	Matchup contribution	Total context contribution	Combined unit ledger
Stephen Curry	Warriors	+1.01	+1.01	+3.40	-0.53	+2.87	+3.88
Draymond Green	Warriors	-2.87	-2.87	+1.83	-0.28	+1.55	-1.32
Jimmy Butler III	Warriors	+4.37	+4.37	+1.33	-0.45	+0.89	+5.26
Moses Moody	Warriors	+1.45	+1.45	+0.26	+0.02	+0.27	+1.72
Will Richard	Warriors	-2.09	-2.09	-1.56	-0.49	-2.05	-4.13
Victor Wembanyama	Spurs	+7.18	-7.18	-2.27	-0.47	-2.74	-9.91
De'Aaron Fox	Spurs	+0.94	-0.94	-1.08	-0.09	-1.17	-2.11
Devin Vassell	Spurs	+3.15	-3.15	-0.68	-0.04	-0.72	-3.87
Stephon Castle	Spurs	+1.34	-1.34	+0.21	-0.09	+0.12	-1.22
Harrison Barnes	Spurs	+0.46	-0.46	-0.04	+0.03	-0.01	-0.47

This is the central interpretive example. Curry's HPM player term is modest in the 2025-26 completed fit, but the frozen context model assigns a large positive composition contribution to his profile in this particular unit. It does not mean Curry alone created all of that environment: the Shapley calculation shares overlapping nonlinear profile effects across Curry, Green, Butler, and the other players in proportion to their average marginal effect.

Rockets Unit Vs Clippers Unit

For Kevin Durant, Alperen Sengun, Jabari Smith Jr., Tari Eason, and Amen Thompson against Brook Lopez, Kawhi Leonard, Kris Dunn, Derrick Jones Jr., and John Collins, the frozen portable edge is (-1.21) and the matchup residual is (+0.61). Eason receives the largest positive Rockets context contribution ((+1.68)), followed by Sengun ((+1.18)); Durant's profile contribution is (+0.23). This is a useful reminder that an excellent player rating and a contextual profile contribution answer different questions.

Knicks Unit Vs Hawks Unit

For Karl-Anthony Towns, OG Anunoby, Josh Hart, Mikal Bridges, and Jalen Brunson against CJ McCollum, Nickeil Alexander-Walker, Onyeka Okongwu, Jalen Johnson, and Dyson Daniels, the frozen total context edge is only (+0.27). Hart ((+2.01)), Towns ((+1.75)), and Brunson ((+1.45)) are the largest positive Knicks context contributors, while Bridges is negative ((-1.19)). The Hawks' largest offsetting context terms are Dyson Daniels ((-1.56)) and Okongwu ((-1.31)).

Interpretation Boundaries

This is a profile-based counterfactual inside HPM's frozen context function. It does not observe a player being literally replaced by an average player.
The synthetic reference unit is only an accounting zero point. It does not replace HPM's empirical, possession-weighted reference field used to define portable composition ratings.
A positive opponent contribution helps the focal unit's edge; a negative one helps the opponent. This sign convention lets all ten rows sum directly.
The values are lineup- and opponent-dependent. A player can have a different context contribution in another five-man unit even with the same HPM player rating.
Low exact-matchup exposure means these selected examples are pedagogical decompositions, not estimates of how those specific units performed on court.

Artifact

Immutable artifact:

artifacts/models/portable_context_attribution_audit/2025-26/

case_summary.parquet contains the ledger totals and realized matchup exposure. player_context_attribution.parquet contains all ten player rows for every case. Regenerate it with the audit guide.