The skill model
Thirteen skills.
Deliberately not one number.
SAGE measures what a player can do, one skill at a time, from what actually happened on the floor. It does not add them up. A shooting percentage, a rebound and a forced turnover are different currencies, and any single number that combined them would be a claim the evidence does not support.
Three questions, asked of every candidate skill in turn. A measurement that cannot answer all three does not get published — and most of them could not. The list below is what survived.
One night, worked through
Alex Caruso in Dallas, 10 December 2024
The league records who guarded whom, and for how long. That night Caruso spent 46 possessions as somebody’s assigned defender — most of them on the two hardest covers in the building.
| Guarding | Possessions | His turnover rate | Expected |
|---|---|---|---|
| Kyrie Irving | 19.2 | 3.3% | 0.64 |
| Luka Dončić | 13.9 | 5.9% | 0.84 |
| Spencer Dinwiddie | 4.5 | 3.1% | 0.14 |
| five others | 8.4 | — | 0.15 |
| Total | 46.0 | — | 1.77 |
An average defender drawing that assignment forces about 1.6 turnovers. Caruso, who was already rated well above average, was expected to force 2.3. He forced none. So the rating went down — from 3.83 to 3.78 forced turnovers per 100 possessions, about a sixth of one standard error.
That is the size the model chooses for itself. It had already watched him defend 17,652 possessions; one quiet night against elite guards is weak evidence against all of it, and moves the number accordingly. A rookie’s first bad night moves his rating far further, because there is nothing behind it yet.
Idea one · Ability
Doing a lot is not the same as doing it well.
The oldest mistake in basketball statistics is reading a rate as a skill. A player with eight rebounds a game is not thereby a better rebounder than one with four — he may simply be the tallest man on a team that misses a lot.
So every dimension models the opportunity first and the ability second, and publishes both. Ball Security asks what a player’s touches, passes and drives should produce before it asks whether he beat that. Rebounding is solved as a five-man problem rather than counted. Drives are judged on whether anything came of them, not on how many he took.
The test is whether a rating survives a transfer. A skill belongs to the player and should travel with him; a number that collapses when he changes team was measuring his old team.
Idea two · Evidence
Every rating is what we knew before tip-off.
A rating for a game never uses that game. It is a running estimate that starts knowing nothing — a player’s first night carries the league average exactly — and is nudged once per game, forever. Nothing is ever recomputed from a season average.
The size of each nudge is decided by the evidence already banked, not by a setting. And uncertainty is carried openly: on the player page every skill shows a range as well as a number, and a wide range means the honest answer is that the model cannot separate him from average yet.
Time counts in calendar days, not games, because ability drifts during an injury or an off-season when nothing is being observed. A player who has not played in two months is rated less certainly than the day he stopped.
Idea three · Worth
A skill has to earn its place twice.
A measurement earns its place here by clearing two bars, not one. It has to belong to the player — travelling with him when he changes team — and it has to show up in the score.
What clears both are the moments that end a possession and can be pinned on one man: a forced turnover, a foul not given, a shot missed at the rim. Forcing your man into a turnover is the strongest of them — it is worth about half a standard deviation in points a team concedes, and one season of it separates the best on-ball defenders from average by roughly twenty points. Those are the four defensive dimensions published here, and they are published because they passed, not because the category needed filling.
The second bar is what makes the first trustworthy. Plenty of defensive measurements clear only the first — how many points the man you guard scores repeats from season to season and travels between teams, and yet barely moves what a team concedes, because the shots simply arrive somewhere else. Holding every candidate to both bars is why the four that survived can be read as defence rather than as style.
What it measures
Thirteen dimensions, each in its own unit, each with its own uncertainty. Click any one for how it is measured and what it cannot see.
Offence
The glass
Defence
What it will not tell you
There is no total. Thirteen skills in thirteen units do not add up, and the weights that would combine them are not something this data can settle. For a single impact number there is AURA, which answers a different question entirely.
These are estimates of ability now, not forecasts. Projection is ORACLE’s job; it reads these ratings as inputs, which is safe precisely because each one was computed without knowing the game it belongs to.
Screen navigation, closeout quality, rotational discipline, communication, gravity. None of these has an event with a name attached to it in this data, and no amount of statistical work conjures one.
Ratings are league-wide by default, and some skills are shaped by position — a guard will rate low on the defensive glass because guards do. The player page carries a toggle to compare a player against his own position instead.
SAGE v1.3, figures as published 30 August 2026. Every rating on this page is a real published state, and the worked example is Alex Caruso’s actual row for that date. See the skills themselves on the player page. For team-level impact rather than skills, see AURA; for what a player will produce, the projection model.