This report introduces a first pass at player event Elo models.

Each player gets two numbers for each event family: - for_team: contribution to the event happening for the player's team - against_team: contribution to preventing that event for the opponent

For points, that gives a clean interpretable 2D view: - Points For Elo = points created for your own team - Points Prevention Elo = points prevented from the other team - Points Net Elo = the sum

Important caveat: these are online event decompositions with actual played minutes, not adjusted impact estimates. They are much more interpretable than the earlier raw Vec2, but they still inherit a lot of team context and are not a replacement for RAPM.

Current defaults: learning_rate=0.08, shrink=0.004.

Main Read

NBA panel: top 50 players by RAPM Overall among players with 1000+ minutes, plotted in the event-Elo points space.

Arizona panel: every Arizona player with a current-season event-Elo row.

What I Learned

• The points split is immediately more interpretable than raw Vec2. You can read offense and defense directly instead of guessing latent axes.
• Arizona looks sensible in this frame: B. Burries, J. Bradley, K. Peat, and I. Kharchenkov all grade as strong two-way point contributors in the current fit.
• The family has real holdout lift on several event types, especially NBA points/rebounds/stocks and NCAAM assists/threes/turnovers.
• The biggest weakness is exactly what the raw leaderboards suggest: the model is still team-context heavy. NBA Points Net Elo only correlates modestly with RAPM Overall, so this is not yet an adjusted all-in player value model.
• In other words: this is already a useful event microscope, but not yet the final answer to 'best player' in the same sense as RAPM.

Holdout

Strict holdout here means: fit on games before the trailing 7-day window, then predict the final 7 days using the actual played-minute mix in those holdout games.

NBA Holdout

Event	Model RMSE	Mean RMSE	Lift	Corr	Team Games
Points	10.555	10.986	0.430	0.431	100
Rebounds	6.214	6.501	0.288	0.503	100
Offensive Rebounds	3.761	3.967	0.205	0.468	100
Defensive Rebounds	5.152	5.331	0.179	0.466	100
Assists	4.606	4.807	0.202	0.427	100
Threes	3.900	4.118	0.218	0.499	100
Stocks	3.853	3.968	0.115	0.302	100
Turnovers	3.613	3.755	0.142	0.402	100

NCAAM Holdout

Event	Model RMSE	Mean RMSE	Lift	Corr	Team Games
Points	15.576	15.438	-0.138	0.392	618
Rebounds	8.750	8.864	0.115	0.391	618
Offensive Rebounds	6.510	5.567	-0.943	0.269	618
Defensive Rebounds	11.502	6.583	-4.919	0.133	618
Assists	6.209	6.500	0.290	0.467	618
Threes	4.508	4.723	0.215	0.493	618
Stocks	4.824	4.950	0.126	0.293	618
Turnovers	4.720	4.863	0.143	0.361	618

Positive Lift means the event model beat a flat holdout mean baseline.

Cross-Model Overlap

League	Qualified	PtsNet vs RAPM	PtsNet vs PM/100	RebNet vs RAPM	TONet vs RAPM
nba	220	0.420	0.770	0.307	0.174
ncaam	345	0.740	0.814	0.563	0.301

Read this as overlap, not as a target. Lower correlation with RAPM is not automatically bad; it can also mean the event family is describing something more specific.

NBA Top 50 by RAPM in Points Space

RAPM Rank	Player	Team	Pts For	Pts Prevent	Pts Net	OReb Net	Stocks Net	RAPM
1	Victor Wembanyama	SAS	0.898	1.610	2.509	-0.150	0.129	4.474
2	Cade Cunningham	DET	0.812	2.851	3.663	1.249	1.008	3.999
3	Kevin Durant	HOU	0.658	0.794	1.451	1.306	0.721	3.681
4	Nikola Jokić	DEN	3.177	-1.183	1.994	0.001	-1.005	3.555
5	Collin Gillespie	PHX	-0.697	0.850	0.153	0.477	0.584	3.305
6	Jaylin Williams	OKC	0.934	1.983	2.917	-0.757	0.424	3.026
7	Austin Reaves	LAL	1.254	-0.224	1.030	-0.646	0.000	2.968
8	Neemias Queta	BOS	1.282	1.638	2.920	0.874	0.628	2.898
9	Oso Ighodaro	PHX	-0.421	0.557	0.136	0.313	0.442	2.874
10	Jalen Duren	DET	0.394	2.325	2.719	0.869	0.805	2.818
11	Michael Porter Jr.	BKN	-1.613	-1.086	-2.699	0.379	-1.013	2.654
12	Svi Mykhailiuk	UTA	-0.136	-1.791	-1.927	0.189	-0.845	2.578
13	Jimmy Butler III	GSW	0.947	0.584	1.531	-0.070	0.496	2.458
14	Luka Dončić	LAL	1.344	-0.211	1.133	-0.663	0.477	2.438
15	Deni Avdija	POR	-0.655	-0.803	-1.458	0.453	-2.037	2.411
16	Mikal Bridges	NYK	1.549	1.519	3.069	1.326	0.240	2.399
17	Tyrese Maxey	PHI	0.223	0.691	0.914	0.009	1.433	2.381
18	Donovan Clingan	POR	-0.204	-0.061	-0.265	0.208	-1.472	2.367
19	Shai Gilgeous-Alexander	OKC	1.155	3.568	4.723	-1.569	1.576	2.274
20	Jose Alvarado	NOP	-0.545	-0.134	-0.680	-0.161	0.248	2.222
21	Bam Adebayo	MIA	-0.095	1.335	1.241	0.526	-0.057	2.182
22	Amen Thompson	HOU	0.319	1.507	1.826	1.952	0.297	2.176
23	Donte DiVincenzo	MIN	0.648	0.796	1.444	-0.008	1.102	2.146
24	Sandro Mamukelashvili	TOR	-0.128	0.684	0.556	0.025	0.344	2.144
25	Dylan Harper	SAS	1.219	1.220	2.439	-0.476	0.422	2.081
26	Wendell Carter Jr.	ORL	-0.890	1.024	0.133	0.101	0.220	2.068
27	Quinten Post	GSW	-0.203	0.690	0.487	-0.236	0.198	2.063
28	Jalen Suggs	ORL	-0.123	1.125	1.002	0.060	0.210	2.063
29	Jaime Jaquez Jr.	MIA	0.184	1.415	1.599	0.015	-0.018	2.050
30	Ajay Mitchell	OKC	1.606	2.801	4.407	-0.911	1.517	2.040
31	Dyson Daniels	ATL	0.291	0.654	0.945	-0.536	0.139	1.944
32	Tim Hardaway Jr.	DEN	2.155	-1.163	0.992	-0.712	-0.850	1.891
33	Noah Clowney	BKN	-1.707	-1.193	-2.900	0.158	-0.864	1.765
34	Keldon Johnson	SAS	0.951	1.582	2.533	0.095	0.388	1.726
35	Isaiah Joe	OKC	0.730	2.162	2.892	-0.785	0.764	1.701
36	Jamal Murray	DEN	3.454	-2.171	1.283	-0.945	-1.165	1.689
37	Chet Holmgren	OKC	1.113	3.562	4.675	-0.964	1.200	1.656
38	Evan Mobley	CLE	0.479	0.681	1.159	0.259	0.353	1.613
39	Tristan da Silva	ORL	-0.124	0.725	0.601	0.058	-0.010	1.609
40	Ausar Thompson	DET	0.460	2.257	2.718	0.612	0.873	1.580
41	OG Anunoby	NYK	1.162	1.243	2.406	1.107	0.138	1.578
42	Jarrett Allen	CLE	1.312	0.600	1.912	-0.034	0.175	1.572
43	Vít Krejčí	ATL	-0.331	-0.321	-0.653	-0.000	-0.174	1.487
44	Norman Powell	MIA	0.141	0.518	0.659	0.193	-0.275	1.431
45	Sidy Cissoko	POR	-0.626	-0.195	-0.821	0.594	-1.172	1.417
46	Rui Hachimura	LAL	1.233	-0.473	0.759	-0.588	0.052	1.407
47	Davion Mitchell	MIA	-0.011	1.384	1.373	-0.221	0.232	1.383
48	Jaden McDaniels	MIN	0.805	0.645	1.450	-0.037	1.048	1.341
49	Myles Turner	MIL	-0.910	-0.937	-1.846	-1.015	-0.640	1.308
50	Pascal Siakam	IND	-1.383	-1.563	-2.945	-0.398	-0.200	1.299

Arizona Roster in Points Space

Team Rank	Player	Pts For	Pts Prevent	Pts Net	OReb Net	DReb Net	Stocks Net	Minutes
1	B. Burries	4.396	2.665	7.061	1.112	3.720	1.053	887.0
2	J. Bradley	4.179	2.632	6.811	1.058	3.795	1.072	911.0
3	K. Peat	3.804	2.660	6.464	0.818	3.747	0.944	732.0
4	I. Kharchenkov	3.920	2.195	6.115	1.000	3.345	0.849	816.0
5	M. Krivas	3.402	2.327	5.729	0.980	3.162	0.817	746.0
6	A. Dell'Orso	3.337	2.109	5.447	0.920	2.761	0.776	648.0
7	T. Awaka	3.127	2.088	5.215	0.774	2.774	0.796	648.0
8	D. Aristode	2.380	1.992	4.372	0.554	2.162	0.656	391.0
9	S. Gueye	0.514	0.406	0.919	0.142	0.357	0.151	78.000
10	E. Nelson	0.526	0.379	0.905	0.158	0.323	0.164	73.000
11	A. Arnold	0.284	0.280	0.564	0.097	0.241	0.072	38.000
12	S. Djopmo	0.220	0.235	0.454	0.074	0.194	0.059	30.000
13	J. Francois	0.090	0.108	0.197	0.025	0.087	0.031	12.000
14	J. Cook	0.096	0.101	0.197	0.031	0.083	0.034	12.000

Path Forward

• Use these event models as the interpretable 2D baseline: points_for and points_against are the clean intermediate win the earlier Vec2 was missing.
• Add projected-minute inputs so holdout predictions stop using realized minute shares.
• Add teammate/opponent adjustment or hierarchical team priors so the raw leaderboards stop inheriting as much team strength.
• Rebuild the nonlinear vector model on top of this decomposition rather than jumping straight to opaque latent axes.

Reproduce

python scripts/generate_player_event_elo_report.py --nba-season 2025-2026 --team-season 2025-2026 --team-name "Arizona"