RealGM

I’ve been working on a draft model designed to predict peak NBA performance using a combination of statistical production, anthropometric measurements, and scouting evaluations. I plan to continue refining the model, but I believe it’s already at a stage where it’s worth sharing. Below, I’ve presented the model’s current top 30 NCAA prospects for the 2026 draft class.

P.S. (Jan 8, 2026) - Model includes all NCAA players ranked on NBAdraft.net or ESPN top 100 big boards with the exception of Alijah Arenas (19 on NBAdraft.net, unranked on ESPN) who has not played an NCAA game.


Methodology (skip if you wish)

To measure peak performance, I use peak DPM. DPM is a hybrid box-score and regression-based metric, similar in spirit to EPM, LEBRON, xRAPM, or ESPN’s now-defunct RPM. I chose DPM for two main reasons. First, in a poll conducted a few years ago, NBA executives voted DPM the best publicly available metric, narrowly edging out EPM. Second, peak DPM data is freely accessible via nbarapm.com, whereas EPM requires a subscription. I also avoided RAPM because its estimates are highly unreliable for players with small minute samples, which is common in this dataset.

The dataset includes all drafted prospects from 2008 through 2021. I rely on the Barttorvik database for college basketball data, as it is both comprehensive and easy to work with, though it only extends back to 2008. I selected 2021 as the cutoff because more recent draftees still have significant room to develop before reaching their peak. While some 2020–21 draftees may still improve meaningfully, extending the cutoff earlier allows for a larger and more stable sample of players who have plausibly reached their peak performance.

For scouting input, I use historical Top 100 draft boards from NBADraft.net, as well as ESPN’s Top 100 boards when available. NBADraft.net provides complete boards going back to 2008, while ESPN boards were accessible for many—but not all—years. I am open to incorporating additional scouting sources, provided they cover multiple seasons within the 2008–2021 range.

Anthropometric data is sourced primarily from the NBA’s official website. When that information is unavailable, I supplement it with data from CraftedNBA or NBADraft.net to at least capture height and weight. If a player’s wingspan is not listed, I assume it to be 3.5 inches greater than height, which is approximately the NBA average. Since no official measurements exist yet for 2026 prospects, current projections rely on their listed height and weight from NBADraft.net, along with the same assumed +3.5-inch wingspan.

Model Inputs

The model currently incorporates the following inputs:

Consensus scouting rank*
Position
Draft age
BPM
Offensive rebound rate
Assist rate
Steal rate
Block rate
Position-adjusted wingspan**
Predicted three-point percentage***

* Consensus scouting rank: The average of a prospect’s ranking across multiple scouting boards (currently NBADraft.net and ESPN). Unranked prospects are assigned a default rank of 125.

** Position-adjusted wingspan: The difference between a player’s wingspan and the average wingspan for their listed position within the dataset. Positive values indicate longer-than-average wingspans; negative values indicate shorter-than-average wingspans.

*** Predicted three-point percentage: Career three-point percentage is estimated via linear regression using free-throw percentage, three-point percentage, and three-point attempt rate as explanatory variables.

Further Notes

L1 regularization was used during development to inform variable selection, but the final model applies no regularization. In testing, regularization did not meaningfully reduce out-of-sample mean squared error. The variables listed above represent the most comprehensive feature set I found that preserved predictive performance without introducing excessive overfitting.

Simpler specifications—such as models excluding anthropometric variables or ignoring block rate for non-bigs—produce similar out-of-sample error, but I prefer this model since it considers the most facets of the game.

P.S. (Jan 8) - Should've mentioned that players who do play a game in the NBA are assigned a peak DPM of -2 which appears to be close to what DPM uses as a default prior is for an undrafted player. Moreover, players whose peak DPM is less than -2 are assigned a peak DPM value of -2.


***********************************************
PROSPECTS (As of Jan 6, 2026)
***********************************************

Rank	Prospect	Position	College	Class	Consensus	Estimated Peak DPM
1	Cameron Boozer	PF	Duke	Fr	3.0	3.48
2	Darryn Peterson	PG/SG	Kansas	Fr	1.5	2.60
3	AJ Dybantsa	SF	BYU	Fr	1.5	2.44
4	Caleb Wilson	PF	North Carolina	Fr	4.5	1.91
5	Kingston Flemings	PG	Houston	Fr	6.0	1.85
6	Yaxel Lendeborg	PF/C	Michigan	Sr	17.0	1.52
7	Hannes Steinbach	PF	Washington	Fr	12.5	0.95
8	Labaron Philon	PG	Alabama	So	14.5	0.91
9	Patrick Ngongba	C	Duke	So	24.5	0.82
10	Nate Ament	SF/PF	Tennessee	Fr	7.0	0.78
11	Jayden Quaintance	PF/C	Kentucky	So	21.5	0.58
12	Zuby Ejiofor	PF/C	St. John's	Sr	34.5	0.56
13	Joseph Tugler	SF/PF	Houston	Jr	59.5	0.54
14	Bennett Stirtz	PG	Iowa	Sr	15.5	0.41
15	Joshua Jefferson	PF	Iowa State	Sr	54.5	0.37
16	Mikel Brown Jr.	PG	Louisville	Fr	4.5	0.36
17	Koa Peat	PF	Arizona	Fr	16.0	0.33
18	JT Toppin	PF/C	Texas Tech	Jr	38.5	0.33
19	Morez Johnson Jr.	PF	Michigan	So	45.0	0.27
20	Miles Byrd	SG	San Diego St.	Jr	32.0	0.27
21	Keaton Wagler	SG	Illinois	Fr	41.0	0.26
22	Matt Able	SG	NC State	Fr	26.5	0.17
23	Tounde Yessoufou	SG	Baylor	Fr	22.0	0.16
24	Thomas Haugh	SF	Florida	Jr	14.0	0.15
25	Tyler Tanner	PG	Vanderbilt	So	97.0	0.15
26	Darius Acuff Jr.	PG/SG	Arkansas	Fr	16.0	0.14
27	Alex Condon	PF/C	Florida	Jr	37.5	0.14
28	Aday Mara	C	Michigan	Jr	59.0	0.13
29	Isaiah Evans	SG/SF	Duke	So	18.5	0.04
30	Dailyn Swain	SF/PF	Texas	Jr	60.0	0.03
31	Meleek Thomas	PG/SG	Arkansas	Fr	21.0	0.02
32	Cameron Carr	SG	Baylor	So	15.0	-0.10
33	Tarris Reed Jr.	C	Connecticut	Sr	67.5	-0.12
34	Braylon Mullins	SG	Connecticut	Fr	14.0	-0.18
35	Johann Grunloh	PF/C	Virginia	Fr	49.0	-0.19
36	Acaden Lewis	PG	Villanova	Fr	110.0	-0.22
37	Henri Veesaar	PF/C	North Carolina	Jr	28.5	-0.26
38	Brayden Burries	PG/SG	Arizona	Fr	33.5	-0.28
39	Eric Reibe	C	Connecticut	Fr	42.0	-0.28
40	Motiejus Krivas	C	Arizona	Jr	61.0	-0.30
41	KJ Lewis	SG	Georgetown	Jr	72.0	-0.30
42	Christian Anderson	PG	Texas Tech	So	52.0	-0.31
43	Dame Sarr	SG/SF	Duke	Fr	33.5	-0.31
44	Nate Bittle	C	Oregon	Sr	72.5	-0.33
45	David Punch	PF	TCU	So	91.5	-0.36
46	Richie Saunders	PG/SG	BYU	Sr	47.0	-0.36
47	Anthony Robinson II	PG/SG	Missouri	Jr	102.5	-0.36
48	Magoon Gwath	PF/C	San Diego St.	So	42.5	-0.46
49	David Mirkovic	PF	Illinois	Fr	88.0	-0.47
50	Zvonimir Ivisic	PF/C	Illinois	Jr	110.0	-0.47
51	Flory Bidunga	PF/C	Kansas	So	72.5	-0.48
52	Alex Karaban	SF/PF	Connecticut	Sr	44.0	-0.55
53	Elyjah Freeman	SG/SF	Auburn	So	70.0	-0.57
54	Amaël L’Etang	C	Dayton	So	69.0	-0.60
55	Chris Cenac Jr	PF/C	Houston	Fr	10.5	-0.61
56	Boogie Fland	PG	Florida	So	57.5	-0.64
57	Braden Smith	PG	Purdue	Sr	80.5	-0.67
58	Ja'Kobi Gillespie	PG	Tennessee	Sr	110.5	-0.69
59	Baba Miller	C	Cincinnati	Sr	80.0	-0.70
60	Nolan Winter	PF/C	Wisconsin	Jr	81.5	-0.74
61	Kwame Evans Jr.	PF	Oregon	Jr	101.5	-0.75
62	Oscar Cluff	PF/C	Purdue	Sr	109.5	-0.77
63	Malik Reneau	SF/PF	Miami FL	Jr	100.0	-0.85
64	Dillon Mitchell	PF	St. John's	Sr	94.5	-0.86
65	Paul McNeil, Jr.	SG	N.C. State	So	92.5	-0.86
66	Ryan Conwell	SG	Louisville	Sr	93.5	-0.88
67	Neoklis Avdalas	SG/SF	Virginia Tech	Fr	24.0	-0.91
68	Collin Chandler	SG	Kentucky	So	67.5	-0.92
69	Emanuel Sharp	PG/SG	Houston	Sr	106.5	-0.94
70	PJ Haggerty	PG	Kansas St.	Jr	82.5	-0.96
71	Jaden Bradley	PG	Arizona	Sr	108.5	-0.97
72	Karter Knox	SF	Arkansas	So	54.0	-0.98
73	Wesley Yates III	SG	Washington	So	64.5	-1.00
74	Ian Jackson	SG	St. John's	So	60.0	-1.01
75	Kylan Boswell	PG/SG	Illinois	Sr	94.0	-1.03
76	Tomislav Ivisic	C	Illinois	Jr	47.0	-1.08
77	Juke Harris	SF	Wake Forest	So	84.0	-1.12
78	Darrion Williams	SF	N.C. State	Sr	83.5	-1.12
79	Robert Wright	PG/SG	BYU	So	108.5	-1.12
80	Moustapha Thiam	C	Cincinnati	So	78.0	-1.14
81	Tahaad Pettiford	PG	Auburn	So	66.0	-1.15
82	Keyshawn Hall	SG/SF	Auburn	Sr	77.5	-1.16
83	Baye Ndongo	PF/C	Georgia Tech	Jr	105.0	-1.16
84	Jalen Haralson	SG/SF	Notre Dame	Fr	97.0	-1.18
85	Solo Ball	SG	Connecticut	Jr	57.5	-1.22
86	Taylor Bol Bowen	SF	Alabama	Jr	110.5	-1.25
87	Bruce Thornton	PG/SG	Ohio St.	Sr	102.5	-1.26
88	Tyrone Riley IV	SF	San Francisco	So	101.5	-1.29
89	Otega Oweh	SG	Kentucky	Sr	107.0	-1.31
90	Malique Ewin	PF/C	Arkansas	Sr	104.0	-1.32
91	Pryce Sandfort	SG/SF	Nebraska	So	106.0	-1.38
92	Mouhamed Sylla	C	Georgia Tech	Fr	95.0	-1.39
93	John Blackwell	SG	Wisconsin	Jr	101.0	-1.40
94	Andrej Stojakovic	SG/SF	Illinois	Jr	60.0	-1.41
95	Jaland Lowe	PG	Kentucky	Jr	112.5	-1.42
96	Milan Momcilovic	PF	Iowa St.	Jr	95.5	-1.44
97	Trevon Brazile	PF	Arkansas	Sr	103.5	-1.44
98	Trey Kaufman-Renn	SF/PF	Purdue	Sr	105.0	-1.48
99	Coen Carr	SF	Michigan St.	Jr	80.5	-1.48
100	Tucker DeVries	SG/SF	Indiana	Sr	95.0	-1.58
101	Tre White	SG	Kansas	Sr	104.5	-1.58
102	D.J. Wagner	PG/SG	Arkansas	Jr	107.5	-1.61
103	Mackenzie Mgbako	SF/PF	Texas A&M	Jr	74.0	-1.64
104	Milos Uzan	PG/SG	Houston	Sr	57.0	-1.67
105	Naithan George	PG/SG	Syracuse	Jr	105.5	-1.67
106	Nick Martinelli	SF	Northwestern	Sr	108.0	-1.69
107	Josh Dix	SG	Creighton	Sr	103.0	-1.74
108	Tyler Harris	SG/SF	Vanderbilt	Jr	87.5	-1.79
109	Jasper Johnson	PG/SG	Kentucky	Fr	99.5	-1.79
110	Felix Okpara	C	Tennessee	Sr	111.5	-1.80
111	Tobi Lawal	PF	Virginia Tech	Sr	77.0	-1.89
112	Joson Sanon	SG	St. John's	So	106.0	-1.96
113	Jaron Pierre Jr.	PG/SG	SMU	Sr	70.5	-2.14
114	Donald Hand	SG	Boston College	Jr	111.0	-2.35

IMPROVEMENTS

1. Use past seasons (not just the current season) in projections. This should help stabilize the projections for non-freshmen who have outlier (good or bad) final seasons. High school / AAU stats could be used for freshmen but I don't have any high-school data outside of recruit rank. (DONE)

2. Position-specific projections. I tried creating position-specific projections, and while I got significantly better predictions for bigs, and decent predictions for guards, the predictions for wings were a disaster.

3. Predict peak DPM for current players using NBA performance, and use that prediction in place of peak DPM. This would account for some of the younger players in the dataset not being at their true peak. Furthermore, it may allow me to use prospects from the most recent drafts in the training set.


I'll expand the prospect list outside the top 30 once I get the data for all the relevant prospects.

The OP has been updated with a complete list of prospects. Any NCAA prospect appearing on either NBADraft.net or ESPN’s Top 100 big boards is included. For players with multiple NCAA seasons, the model uses multi-year data. All statistics reflect games played through January 6, 2026.

MORE TECHNICAL DETAILS

I’m largely satisfied with the current version of the model and do not expect to make further changes for the remainder of the season. After additional tuning to reduce out-of-sample prediction error, I made a few adjustments:

Removed variables (inclusion of these variables consistently made out-of-sample prediction error worse):

Wingspan
Predicted 3-point percentage

Added variables:

Rim frequency – percentage of a player’s 2-point shots taken at the rim
Dunk rate – made dunks per minute played

MODEL PERFORMANCE

To evaluate performance, I randomly subsampled 20% of the data as test data (with the remaining 80% used for training) and repeated this process 10,000 times. Below is a summary of the results (MSE = mean squared error):

Best Linear Model – the current model
NBA Draft Order – vanilla linear regression using only draft position to predict peak DPM
Scouting Only – vanilla linear regression using only consensus scouting rank to predict peak DPM

Model	MSE	R^2
Best Linear Model	2.0383	0.2647
NBA Draft Order	2.0887	0.2451
Scouting Only	2.242	0.1944

The results show that actual NBA draft position is a significantly better predictor of success than pre-draft big boards. It’s unclear how much of this advantage comes from NBA teams being better evaluators versus teams investing more heavily in the development of highly drafted players. I am pleased that the current model meaningfully outperforms NBA draft position.

Is there collinearity with the block/assist/steal/rebounding rates and BPM? Since BPM includes box-score stats to calculate I would imagine there is (I am not sure if it matters much or the solution here - perhaps it is to have a draft age-BPM ranking, and a draft age-box score metric ranking).

FWIW I really like this : )

babyjax13 wrote:Is there collinearity with the block/assist/steal/rebounding rates and BPM? Since BPM includes box-score stats to calculate I would imagine there is (I am not sure if it matters much or the solution here - perhaps it is to have a draft age-BPM ranking, and a draft age-box score metric ranking).

FWIW I really like this : )

Happy your interested!

There are collinearity issues with BPM. In my datababse, BPM is highly correlated with TS (r=0.617), EFG (r=0.545), and to a lesser extent TOV (r=-0.402), but I don't use any of those stats in the model. It's worth noting that BPM is more strongly correlated to scouting rank than any statistic I use in the model - despite some of those stats being directly used in BPM. The correlation between scouting rank and BPM is -0.277, while STL% has the next strongest correlation with BPM at 0.233.

That being said, if, for example, you include both BPM and AST% in your model, in a way, you are including AST% twice. It's why some variables (3PT%, from my testing) can paradoxically end up with negative coefficients. It's not that having a higher 3PT% is bad - it's that BPM is already capturing 3PT%, and that the negative coefficient spit out by the model is the model's way of saying that BPM values 3PT% too high. As a tangent, the only coefficient in my final model whose sign is opposite what you'd expect is dunk rate (negative coefficient). Still, high dunk rate tend to have better model predictions because I also include frequency of shots at the rim as a variable and high dunk guys tend to be high rim rate guys. I suspect the model is trying to fade the garbage men with hyper-efficient shooting percentages who get a large percentage of their offense through assisted dunks. As a tangent, I suspect similar reasons as to why anthropometric data did not improve performance - scouting rank already takes into account that information. This is supported by the fact that anthropometric data significantly improves the performance of a stats-only model.

Ultimately, I'm not sure how big a problem collinearity actually is. Collinearity increases the uncertainty of the estimated coefficients - it does not inherently bias the results one way or another. The estimated coefficients for BPM (and the other variables) appear to be somewhat stable across different training sets, so I don't think collinearity is wreaking havoc on the model. At the end of the day, short of doing something like PCA, there's going to be some collinearity between the inputs. I tested models that did not use BPM, but they had significantly worse out-of-sample prediction error, which is why BPM ended up making it into the model.

greenOakX wrote:

babyjax13 wrote:Is there collinearity with the block/assist/steal/rebounding rates and BPM? Since BPM includes box-score stats to calculate I would imagine there is (I am not sure if it matters much or the solution here - perhaps it is to have a draft age-BPM ranking, and a draft age-box score metric ranking).

FWIW I really like this : )

Happy your interested!

There are collinearity issues with BPM. In my datababse, BPM is highly correlated with TS (r=0.617), EFG (r=0.545), and to a lesser extent TOV (r=-0.402), but I don't use any of those stats in the model. It's worth noting that BPM is more strongly correlated to scouting rank than any statistic I use in the model - despite some of those stats being directly used in BPM. The correlation between scouting rank and BPM is -0.277, while STL% has the next strongest correlation with BPM at 0.233.

That being said, if, for example, you include both BPM and AST% in your model, in a way, you are including AST% twice. It's why some variables (3PT%, especially from what I found) paradoxically end up with negative coefficients. It's not that having a higher 3PT% is bad - it's that BPM is already capturing 3PT%, and that the negative coefficient spit out by the model is the model's way of saying that BPM values 3PT% too high. As a tangent, the only coefficient in my final model whose sign is opposite what you'd expect is dunk rate (negative coefficient). Still, high dunk rate tend to have better model predictions because I also include frequency of shots at the rim as a variable and high dunk guys tend to be high rim rate guys. I suspect the model is trying to fade the garbage men with hyper-efficient shooting percentages who get a large percentage of their offense through assisted dunks.

I'm not sure how big a problem collinearity actually is. Collinearity increases the uncertainty of the estimated coefficients - it does not inherently bias the results one way or another. The estimated coefficients for BPM (and the other variables) appear to be somewhat stable across different training sets, so I don't think collinearity is wreaking havoc on the model. At the end of the day, short of doing something like PCA, there's going to be some collinearity between the inputs. I did test models that did not use BPM, but they had significantly worse out-of-sample prediction error which is why I ended up including it in the model.

Interesting to read! FWIW I did a first watch on Keaton Wagler and a rewatch of Ejiofor because of this!

Yeah, the model really loves Zuby Ejiofor. As for Keaton Wagler, you can expect him to be a lot higher when I update this model. My guess is he'll be top 10.