How it works

I taught a model to study every college prospect since 2009: how they scored, passed, defended, and rebounded, how young they were, how big they were, and who they did it against. Then I watched what those players actually became in the NBA: stars, starters, role players, or guys who washed out.

When a new prospect comes along, the model finds what history says about players with his profile. It never gives one answer. It gives chances: maybe 60% he becomes a starter, 12% he becomes a star, 5% he never sticks. Nobody knows for sure which one happens. The model doesn't pretend to either.

I tested it the hard way. For every draft from 2009 to 2021, I hid that class from the model, asked it to grade those players using only what was knowable before draft night, then checked its answers against real careers.

On the average pick, NBA teams beat the model. They should. They have workouts, medicals, interviews, and intel I don't. But when the model disagreed loudly with where a player was drafted, it was right far more often than chance. Its favorite overlooked players outplayed their draft slots badly, and the players it liked least underplayed theirs. The lesson: don't use this to re-rank the whole board. Use it to find the players worth a second look.

The market is the draft's collective opinion: where NBA front offices actually pick each player, and where public mock drafts and big boards rank him before the night. When I say the market wins on the average pick, I mean teams' real choices predicted careers better than the model did. When the site shows a pick number or a consensus rank next to a player, that is the market's answer sitting beside the model's, so you can see exactly where they disagree. The model never reads any of it.

The same things a good scout checks on the stat sheet, weighted by what has actually predicted careers. Age matters a lot: a 19-year-old and a 23-year-old putting up the same numbers are not the same prospect. Free-throw shooting says more about a future NBA jumper than college three-point percentage does. Steals separate the ones who stick in the league from the ones who don't. Blocks separate ordinary big men from special ones. And production against real competition beats production against nobody.

The full list of factors, in plain English:

• Age and body. Age on draft night, height, wingspan, standing reach, vertical leap, sprint and agility times from the combine.
• Shooting. Free-throw percentage, twos and threes, how often he shoots from deep, how much of his scoring comes at the rim vs the mid-range, dunk rate.
• Production and efficiency. Scoring efficiency, offensive and defensive ratings, overall impact numbers, all adjusted for the level of competition.
• Playmaking and ball security. Assist rate, turnover rate, assist-to-turnover ratio.
• Defense and rebounding. Steal rate, block rate, rebounding on both ends.
• Track record and trajectory. Total minutes and games, years in college, whether he improved year over year.
• Context. Recruiting rank out of high school, power-conference or not, position.

Just as important is what it ignores: where the player was drafted, mock drafts, and big boards. You will see those numbers all over this site, next to the model's, and that is the point. They are the answer key I grade the model against, never its inputs. You cannot grade the room's opinion if you copied off the room.

Film. Medicals. Character. Work ethic. How a guy handles coaching. The model knows none of that, which is why the lottery is where teams beat it worst, and why every player page shows a range instead of a single number. When it says a player has a 50% star chance with a range of 26 to 80, it is telling you it has real uncertainty. Believe the range.

It also only covers players from Division 1 college basketball. International prospects show market prices only, with a badge saying so.

One more honesty rule about small samples. If a player's last college season was only a handful of games, say an injury cut it short, the model does not pretend those games are his resume, and it does not throw them away either. It blends them with his last full season, each game counting at its real weight, and the site labels the result so you know which seasons the score is built on. A player with no full season at all gets no score, because grading almost nothing would be guessing, and the market price is shown instead.

This is the part built for people who watch the games. Write what you saw: "the jumper is real but he floats on defense." The system reads your note, scores it against a fixed checklist of skills, and nudges the player's chances accordingly. Good news nudges them up. Concerns nudge them down.

One rule keeps everyone honest: a note is evidence, never a veto. No single note, however glowing, can turn a 3% star chance into a 30% one. Your eyes and the numbers each get a vote. Neither gets to overrule the other completely.

Notes also work on international prospects the model cannot score. There, your note updates the market's expectation for his draft position (shown as a dashed bar) instead of a model prior; the model stays silent. And once you have saved notes, the board can sort by your own valuations: your big board, with the model as the reference underneath.

Every NBA team gets a pick, in order, on draft night. Picking 9th means 8 players are already gone by the time it's your turn. The war room lets you stand at any pick, 1 through 60, and see who realistically survives that long. I ran the draft ten thousand times, letting players rise and slide the way they actually do on draft night, including the occasional big fall, and used that to estimate the odds each player is still there when your pick comes up. That is the conversation a front office wants to have before draft week, not during it.

The full technical write-up (the model, the tests, the exact definitions, and every design decision with its reasoning) lives in the project repository, along with everything needed to reproduce every number on this site.