Frozen Preseason Leaderboard

This leaderboard evaluates models as true preseason forecasts. Every player value is frozen before the target season begins. Target-season lineups and exposure are supplied by an oracle, but no target-season score, possession outcome, fitted player adjustment, or playoff result can change the model.

This is distinct from the in-season Leaderboard, where models fit the first 1,044 regular-season games before predicting the final 186.

Information Boundary

The initial baseline predicts 2025-26 from the completed 2024-25 regular-only forward RAPM state:

Component	Frozen source
Player values	2024-25 regular-season forward RAPM
Cold starts	Zero RAPM
Offense-margin mean	2024-25 regular season
Home-court effect	Recovered from the completed 2024-25 RAPM state
Lineups and exposure	Realized 2025-26 oracle allocation
Player refit	None

Prior-season playoffs are intentionally excluded from this first baseline.

For eligible target-season possession \(i\),

\[ \widehat y_i = \overline y_{2024-25} + \frac{\sum_{p\in O_i}r_{p,2024-25} - \sum_{p\in D_i}r_{p,2024-25}}{200} + s_i\frac{h_{2024-25}}{200}, \]

where \(s_i=+1\) for home offense and \(-1\) for away offense. The factor 200 uses the same one-number RAPM-to-possession conversion as the in-season Leaderboard.

Frozen Aging Candidate

The first candidate keeps the same scoring equation, league mean, home-court term, target rows, and oracle lineup exposure. It replaces only the frozen player-prior vector with the forward aging ridge model trained through 2024-25. That model uses preseason age, NBA experience, prior RAPM and exposure, draft profile, height, weight, and age-by-profile interactions. It produces the same 582-player 2025-26 coverage as the lagged-RAPM baseline and has no 2025-26 outcome fields.

Frozen Offense/Defense Candidate

The O/D candidate uses two weighted offensive-rating rows per lineup stint: the offense lineup enters offensive columns and the opponent lineup enters defensive columns. It fits forward regular-only O/D states from 1996-97 through 2024-25, then freezes the completed 2024-25 state before scoring 2025-26. See Offense/Defense RAPM for the exact equations and identification convention.

Possession And Game Results

Regular season and playoffs are evaluated separately. Possession metrics use only possessions with one reconstructed lineup, matching the existing neural evaluation boundary. Game-margin RMSE aggregates those eligible possessions in the home-team frame.

Model	Cohort	Games	Possessions	Possession RMSE	Possession MAE	Game-margin RMSE	Possession skill vs frozen mean	Game skill vs frozen mean
Frozen lagged RAPM	Regular season	1,230	218,810	1.199000	1.142154	14.8894	0.0805%	11.5905%
Frozen aging prior	Regular season	1,230	218,810	1.199062	1.142736	15.0203	0.0702%	10.0297%
Frozen O/D RAPM	Regular season	1,230	218,810	1.198853	1.142664	14.8901	0.1092%	11.5692%
Frozen lagged RAPM	Playoffs	85	14,253	1.192895	1.136163	17.5409	-0.0774%	-9.6446%
Frozen aging prior	Playoffs	85	14,253	1.192332	1.136455	16.4946	0.0170%	3.0460%
Frozen O/D RAPM	Playoffs	85	14,253	1.194642	1.139203	17.8037	-0.3924%	-12.9804%

Bolding marks the better value within each cohort and metric. Future frozen priors must use these exact cohorts.

Team Net Rating

Team net rating uses all regular-season RAPM stints, including allocated multi-lineup possessions. For stint (s), the frozen lineup prediction is converted to a predicted margin using the stint's oracle possession exposure; team margins are then summed and divided by team possessions.

Model	Teams	Net-rating RMSE	Net-rating MAE	Pearson correlation	Spearman correlation
Frozen lagged RAPM	30	4.8538	3.9606	0.6113	0.5493
Frozen aging prior	30	5.0366	4.2395	0.6484	0.6111
Frozen O/D RAPM	30	4.8740	3.9380	0.6113	0.5626

Team Win Totals

The primary win estimate is called Pythagorean wins in this project. It is a forward-safe historical mapping from predicted team net rating to expected win percentage, rather than the traditional points-for/points-against Pythagorean exponent. The mapping is fit by game-weighted least squares on 862 regular-season team-seasons from 1996-97 through 2024-25:

\[ \widehat{\operatorname{WinPct}}_t = \operatorname{clip}\left( 0.499583 + 0.030250\,\widehat{\operatorname{NetRtg}}_t, 0, 1 \right). \]

For an 82-game season, the un-clipped form is approximately

\[ \widehat{\operatorname{Wins}}_t = 40.97 + 2.4805\,\widehat{\operatorname{NetRtg}}_t. \]

The historical calibration's in-sample team win-total RMSE is 2.7446 wins. That value measures only the NetRtg-to-wins relationship; the leaderboard error below also includes error in the preseason NetRtg prediction itself.

Model	Teams	Win-total RMSE	Win-total MAE	Win-percentage RMSE	Spearman correlation
Frozen lagged RAPM	30	10.7006	8.9333	0.1305	0.6238
Frozen aging prior	30	10.9632	9.6337	0.1337	0.6394
Frozen O/D RAPM	30	10.9640	8.9734	0.1337	0.6078

As a diagnostic, the artifact also retains the raw count obtained by awarding each game to the team with the positive predicted margin. That deterministic rule has 14.5258 RMSE and 10.7333 MAE, confirming that it turns small predicted edges into unrealistically extreme records. The raw count conserves exactly 1,230 league wins. Independently calibrated Pythagorean expectations are not normalized to the target schedule after fitting: they total 1,230.9 wins for the lagged baseline and 1,229.8 for the aging prior.

The age/draft/physical profile does capture part of the young-team signal. For example, it moves San Antonio from 32.3 to 39.6 Pythagorean wins, but the actual result was 62 wins. A smooth historical player-development prior cannot forecast the largest discontinuous breakouts.

Predicted Standings And Actual Results

The table is sorted by Pythagorean expected wins. Ranks are league-wide rather than conference-specific because conference is not part of the current team-season data contract.

Pythagorean rank	Team	Pythagorean wins	Predicted NetRtg	Actual rank	Actual W-L	Actual NetRtg	Win error
1	OKC	65.6	+9.92	1	64-18	+11.11	+1.6
2	CLE	59.8	+7.60	8	52-30	+4.11	+7.8
3	NYK	56.1	+6.11	6	53-29	+6.50	+3.1
4	LAC	55.5	+5.86	18	42-40	+1.17	+13.5
5	GSW	55.0	+5.66	20	37-45	-0.56	+18.0
6	DEN	49.5	+3.43	5	54-28	+5.15	-4.5
7	LAL	48.3	+2.96	6	53-29	+1.78	-4.7
8	HOU	47.0	+2.42	8	52-30	+5.36	-5.0
9	MIN	46.6	+2.28	10	49-33	+3.32	-2.4
10	BOS	46.3	+2.16	4	56-26	+8.10	-9.7
11	DET	45.8	+1.95	3	60-22	+8.18	-14.2
12	ORL	41.1	+0.05	13	45-37	+0.63	-3.9
13	POR	40.7	-0.09	18	42-40	-0.29	-1.3
14	ATL	40.1	-0.33	11	46-36	+2.37	-5.9
15	MIL	39.2	-0.73	21	32-50	-6.34	+7.2
16	IND	37.8	-1.29	29	19-63	-7.88	+18.8
17	PHX	37.4	-1.43	13	45-37	+1.50	-7.6
18	CHI	37.1	-1.56	22	31-51	-5.07	+6.1
19	TOR	37.0	-1.61	11	46-36	+2.86	-9.0
20	PHI	36.0	-2.01	13	45-37	-0.18	-9.0
21	MEM	35.3	-2.28	25	25-57	-5.94	+10.3
22	SAC	32.9	-3.25	26	22-60	-10.04	+10.9
23	MIA	32.7	-3.32	17	43-39	+2.25	-10.3
24	SAS	32.3	-3.49	2	62-20	+8.29	-29.7
25	NOP	32.1	-3.58	23	26-56	-4.45	+6.1
26	DAL	31.0	-4.02	23	26-56	-5.37	+5.0
27	UTA	29.8	-4.50	26	22-60	-8.15	+7.8
28	BKN	28.7	-4.93	28	20-62	-10.28	+8.7
29	CHA	27.5	-5.42	16	44-38	+4.97	-16.5
30	WAS	26.6	-5.80	30	17-65	-11.76	+9.6

Artifact

The promoted baseline run is frozen-lagged-prior-2025-26-20260805T011238Z-9ac7c011 under artifacts/models/frozen_prior_evaluation/2025-26/. It contains player priors, source-state declarations, possession and game predictions, team net-rating and win tables, the historical Pythagorean calibration panel, all metric tables, file hashes, and an MLflow index.

The evaluated age/draft/physical candidate is frozen-aging-prior-2025-26-20260805T013515Z-2fb4c418 in the same directory. Its source state records both the 2024-25 reference lagged-RAPM run used for the mean/home-court terms and the 2025-26 aging-prior artifact.

The O/D candidate is frozen-offense-defense-rapm-2025-26-20260805T050511Z-62b718bd under artifacts/models/frozen_offense_defense_rapm/2025-26/.