Hoop DossierWhere our numbers come from
Our projections are a trial: a model we are building and testing in the open, and improving as the season goes. Here is what they are, what they weigh, how we test them and what they cannot know. They stay hidden until you switch on Prediction tools, and every one is marked Experimental.
What a projection is
A projection is the middle of a range, not a call on one night. We play each game out many times; a player projected for 24.6 pts will often score 18 pts or 31 pts, and we show that range too. A chance like "56% to go over" means that in nights like this one, we think he goes over a bit more than half the time.
What goes into it
The model brings together the things a good researcher checks before a game: his recent production, his role and expected minutes, who is available on both teams, the matchup, and the game context such as the spread and the schedule. How much each one counts, and how they combine, is our own work and stays private.
On each game page (with Prediction tools on), tap a projection to see which of these moved it up or down.
The screener is context
The Screener shows how a player's numbers moved in situations before (a teammate out, a back-to-back, a fast opponent). We show those splits next to projections so you can judge for yourself; a split from a handful of games is mostly noise, and the screener says so.
Built to be cautious
Sportsbook prices are very good, and a new model that disagrees with them sharply is usually wrong. So ours is deliberately cautious: it only earns more say as its tested record improves, by a rule we fixed before the season.
What it does not know
Late scratches after our last update, a coach's plan, a player's mood, a new role after only a couple of games, anything not in box scores, injury reports and the news we collect. When something is thin or uncertain, the projection says so in a "Not known" line.
Game projections
For each game we project both scores, the winner's chance and, on the game page, what moved the number and what could still change it, in directions only. It weighs who is expected to play on each side against that team's usual rotation, how the two teams have been playing (their results, adjusted for who they played) and home court. It does not yet weigh the schedule spot (rest, back-to-backs, travel); the game page says so when it applies. It never reads betting lines: the sportsbook is what we test it against, not an input.
How it is tested is below: every game, against the final score and the closing line, with the win chances checked for honesty. A new version starts its own record; the older versions' records stay up.
Game projections: the record
Every finished game with a pregame projection is scored (0 finished games had none): the projection we published last before tip-off (the frozen number) against the final score, and against the sportsbook's closing line on the same game. A miss is how far the projected margin or total was from what happened, in points. "Win chance closer" counts the games where our win chance was nearer the result than the closing odds' (margin removed). Nothing is recomputed after a game.
Only 29 games so far: far too few to judge anything. Averages over a handful of games swing wildly; treat this as a live notebook, not a verdict. We will say when the sample is large enough (about 30 games) to mean something.
| Version | Games scored | Winners picked | Margin miss ours · closing line | Total miss ours · closing line | Win chance closer than the closing odds |
|---|---|---|---|---|---|
| v1since Oct 20 | 16 | 9 of 16 | 15.7 ptsline 13.8 pts | 13.5 ptsline 14.7 pts | 5 of 16 |
| v1.1current · since Oct 22 | 29 | 20 of 29 | 10.2 ptsline 9.2 pts | 15.7 ptsline 15.4 pts | 14 of 29 |
Is the win chance honest? (v1.1)
- When we gave the home team 10% to 20%: they won 0 of 2 (we said 19% on average)
- When we gave the home team 20% to 30%: they won 0 of 2 (we said 24% on average)
- When we gave the home team 30% to 40%: they won 0 of 2 (we said 37% on average)
- When we gave the home team 40% to 50%: they won 1 of 4 (we said 46% on average)
- When we gave the home team 50% to 60%: they won 5 of 9 (we said 56% on average)
- When we gave the home team 60% to 70%: they won 2 of 6 (we said 65% on average)
- When we gave the home team 70% to 80%: they won 1 of 1 (we said 73% on average)
- When we gave the home team 80% to 90%: they won 3 of 3 (we said 82% on average)
A win chance is honest when, over many games, teams given 60-70% win about 65% of them. With few games each bucket is mostly noise.
Averages are of absolute misses. Closing line: Pinnacle with the margin removed when we have it, otherwise the DraftKings line captured just before tip. Game pages with Prediction tools on show each game's report card.
What Experimental means
Every model number is Experimental until it passes a test written down in advance: a long run of settled simulated bets over many game days, beating the closing line, and a profit. Until then it is a trial. Nothing here is betting advice.