Explainer

How our NBA predictions work

Every prediction on Drop Coverage is minutes times production: how long we expect a player to be on the court tonight, multiplied by what he produces per minute, adjusted for the opponent, then given a range the result should land inside 80% of the time.

Minutes come first

Most of what changes from night to night is playing time, so that's where the model starts. A team plays 240 minutes a game (a little more on average, once overtime is counted), and those minutes are shared out. When a player is ruled out, his minutes are shared among the teammates still available. Once teams confirm their starting fives near tip-off, more of them go to the teammates who have absorbed that player's minutes before.

On top of each player's usual role, the model adjusts for the situations that reliably move minutes:

Then production per minute

Each player's points, rebounds, assists and threes per minute come from his recent games, with older games counting for less. On the second night of a back-to-back, assists and threes per minute dip slightly. Rookies and anyone with fewer than ten NBA games get an early estimate that blends their own games with a typical player at their position, with wider ranges, and they're kept out of edges until they've played more.

Then the opponent

Some defences give up more to guards, others to big men. For each opponent we measure what it allows to each position against the league average, with recent games weighted more heavily, and apply that at half strength. Tested on last season, full strength helped late in the season and hurt early, when the numbers are thin; half strength held up in both.

Then a range

A single number hides how uncertain a night is, so every prediction carries a range the result should land inside 80% of the time. The range isn't printed on the board; it's what turns a prediction and a line into a probability, and the track record reports how often results actually landed inside it. What the ranges mean explains how they're built.

What isn't in the prediction

Some things look like they should matter and didn't survive testing. When a star sits, the board shows how each teammate has done without him ("Jokić out: +3.1 without him"), but that shift isn't added to the prediction. Tested across last season, it mostly double-counts minutes the model already hands out. Home and away made almost no difference once minutes were right. We show these as information and leave the number alone until the evidence says otherwise.

Tested before it's used

Every piece above had to make predictions more accurate on games it hadn't seen (last season, split into early and late halves) before it went in. Ideas that looked good but made the numbers worse were left out. Even so, most of a miss is the night itself: with a player's exact minutes known in advance, points predictions would still miss by about four points a game on average.

When the numbers update

The main build runs early each morning Sydney time, after the afternoon injury report in the US. The numbers are rerun after each later injury report, 90 minutes before tip-off and once more just before it. When a prediction moves, the board shows the earlier number and why ("22.4 → 25.1, Jokić out").

What's free and what's for members

Until 30 November everything is open. From 1 December the predictions, probabilities, edges, matchups and with/without numbers are for members; who plays, the bookmaker line and how often each player has cleared it in his last five and ten games stay free, and one game a night is fully open as the free game of the day. Game pages, game trends, team pages, Just the Tip and the track record stay free.