Tested on the 2023-2026 drafts, which it never trained on, 65% of the players on the model's 254-player board were drafted, against 32% for a list of the top statistical producers of the same size, and its board held 95 players who went on to be top-100 picks, against 36. Scored in September, after four weeks of games, 54% of its board was still drafted. For every draft-eligible FBS player-season since 2014 (defenders since 2016, when box-score defense starts), the model learns three things: will he be drafted in one of the next two drafts, will he go in the top 100, and will he go in the first round. Two drafts, not one, so a junior who returns to school is not counted as a miss. The board is ordered by expected draft value, measured in the NFL starts a pick in that slot has historically produced in a player's first four seasons (pick 1 is worth about 33 starts, pick 32 about 26, pick 100 about 11, pick 250 about 3). The inputs follow the published prospect models: Football Outsiders' Playmaker Score (yards per team pass attempt, peak-season market share, early declares), BackCAST (a back's share of team rushing and his receiving), QBASE (completion rate and adjusted yards per attempt with years as a starter) and SackSEER (sack and pass-breakup rates), plus breakout age, the college dominator rating, recruiting star ratings and conference strength. The model is trained on every draft-eligible player, not just the ones who were drafted, so it learns who gets passed over too. What it cannot see: game film, the combine and pro days, medicals and interviews. That is why ESPN's scouts still order the players who get drafted far better (rank correlation with the actual draft order 0.83 for ESPN's board, 0.42 for the model, on the same players), and why offensive linemen are its weakest group: college box scores record nothing an offensive lineman does. Early in a season the board leans on last season's games until this season has enough of its own.
Backtest on drafts the model never saw
Measure
Prospect model
Top producers
Share of a 254-player board that was drafted
65%
32%
Players on that board who became top-100 picks
95
36
Share drafted, scored after week 4
54%
28%
Ranking accuracy, drafted vs not (AUC)
0.952
0.785
Accuracy by position (AUC, drafted vs not)
Position
Prospect model
Top producers
Drafted per year
QB
0.968
0.942
21
RB
0.959
0.912
39
WR
0.961
0.944
60
TE
0.961
0.942
28
OL
0.854
0.5
77
DL
0.954
0.912
82
LB
0.948
0.896
55
DB
0.933
0.85
96
ST
0.933
0.848
8
What the model looks at
Factor
What it measures
Production & role
Per-game volume and, more importantly, market share: his share of the team's receiving yards and touchdowns (dominator rating), yards per team pass attempt, share of carries, share of the team's tackles for loss, sacks and pass breakups, how his latest 13 games compare with the 13 before, and his best single season.
Efficiency
Completion rate and adjusted yards per attempt for passers, yards per carry (shrunk toward average on small samples), yards per catch, and CFBD's predicted points added per play.
Size
Listed height, weight and build, measured against other players at his position.
Recruiting pedigree
The 247Sports composite star rating, rating and national rank he carried out of high school or junior college.
Age & experience
Years since his recruiting class (the best age proxy college data has: an early declare is younger), when he broke out, career games and how many schools he has played for.
Program & competition
Power-conference schedule, the roster's talent composite, last season's SP+ rating and the unit around him: line yards, stuff rate and sack rate on both sides of the ball. For offensive linemen, whose box score is empty, this is most of the signal.
Position
How often his position group gets drafted at all. Prospect pages compare each player with a typical draft-eligible player at his own position, so this one only shows here.