This season
Sports
Every game we called before it started, graded when it ended.
9of 15
Our pick won 60% of the games we called.
Our own chances expected about 9.2 right. A 60% pick still loses 4 times in 10.
Every game since Sun, Oct 4, called before it started. Preseason is graded on its own, below.
By league
Preseason games are graded on their own, not in the record above. Starters sit, so they sit close to a coin flip: tested on the last ten preseasons, our preseason model picked 57% of games right.
Us vs the market
Once 100 graded games have a market price to compare against (15 so far), this shows our picks next to the betting market's on the same games at the same moment, whichever comes out ahead.
The edge, graded
When our ratings and the betting market disagree by 10 points or more, the game goes on the Edge page before it starts and is graded here after it ends. None have been graded yet. Preseason games don't count here.
Our ratings are our read before any market price: team strength, injuries, rest and home edge. Before launch we checked 2,400 past NBA games: when the two disagreed this much, results landed close to the market's number. This is where we find out, in public, whether ours can do better.
Confident calls vs close ones
Every pick
Thu, Oct 8
Wed, Oct 7
Tue, Oct 6
Mon, Oct 5
Sun, Oct 4
How we grade
For each game we take the last number the public could see before it started, and check it against the final score. Our pick is the side we gave the better chance.
We don't take our own word for it: every 15 minutes we read the numbers off this site, as anyone would see them, and keep that log. A game counts only with a number the live site showed before kickoff. A number that never made it to the site can't be graded. (The games of Oct 4, before the log existed, were graded on the number our model saved before kickoff.)
Every game we called counts, right or wrong. Once a game is graded it is never changed or removed. Ties aren't graded. NBA preseason games are graded on their own and marked as preseason.
In preseason, starters play about half the minutes, so we halve the expected margin and widen the range, and a team on the second night of a back-to-back gives up about 4.7 points (coaches rest their starters). We tested this on 676 preseason games from the last ten preseasons, each predicted only from what was known before it: it picked 57% of them right.
On Oct 5 we voided three preseason games we had graded by mistake. Their numbers came from a test run that was never on the site before the games started, so they were not real calls. We fixed the model the same day, and the preseason record starts again with the games of Oct 5.
Brier score so far: 0.216. It is the average squared gap between our number and what happened: 0 is perfect, and saying 50% every time scores 0.250.
Sources
Final scores from ESPN's public scoreboard. Our numbers from our own published history, timestamped when we post them.