Method note
Why defender rankings built on matchup data mislead
A points-allowed-per-possession leaderboard mostly ranks defenders by where on the floor they are assigned to guard, not by how well they guard. Defenders charged with shots at the rim look bad and defenders charged with jump shots look good, because rim shots go in more often. Across 338 NBA defenders, the ones the metric rates worst allowed 50.8% shooting and the ones it rates best allowed 42.9%, a gap that tracks shot location rather than skill.
We built this leaderboard for Edge, found the flaw in it, and did not publish it. What follows is the check that killed it, in enough detail that you can run it yourself or apply it to someone else's version. Figures measured 27 July 2026.
The metric, and why it looks reasonable
The construction is the obvious one. For every possession a defender was matched against a scorer, take the points that scorer produced. Compare the total to what you would expect given that scorer's own rate against everyone else. A defender who allows fewer points than expected has, apparently, suppressed scoring.
It adjusts for opponent quality, which is the objection most people raise first, and it produces a clean single number per defender. It is also the construction behind most public defender leaderboards you will find.
Our version put Matisse Thybulle at the top, 60.0% below expectation. That is a believable result, and a believable result at the top is exactly what stops people checking the bottom.
The bottom of the list is where it falls apart
The defender the metric rated worst in our data was Donovan Clingan, at 91.1% more points allowed than expected. Ranked immediately around him: Luka Garza, Thomas Bryant, Luke Kornet, Joan Beringer. Every one of those is a center, and several are regarded as capable rim protectors.
That is not a list of the league's worst defenders. It is a list of players who stand near the basket. NBA matchup tracking attributes a possession to the defender nearest the scorer, so centers absorb layups, dunks, and post finishes, which are the highest-percentage shots in basketball. Perimeter defenders absorb jump shots, which are not. The metric charges each defender for the shots his position hands him.
Put plainly: a rim protector is penalised for protecting the rim, because the shots taken there go in more often than the shots taken anywhere else, no matter who is standing in the way.
The test that settles it
Naming positions is not proof, so we did not rely on it. Instead, split all 338 qualifying defenders into five equal bands by the metric, then ask a question the metric never sees: what field-goal percentage did each band allow? If the metric measured defensive skill, shooting percentage would move around within the bands. If it measures shot location, the two will move together.
They move together, across every band, without a single reversal.
| Band | Points allowed vs expected | FG% allowed |
|---|---|---|
| 1st· Rated best | -28.2% | 42.9% |
| 2nd | -12.2% | 45.5% |
| 3rd· Middle | +0.3% | 48.4% |
| 4th | +15.3% | 49.9% |
| 5th· Rated worst | +40.9% | 50.8% |
Nearly eight percentage points of shooting separate the band the metric calls best from the band it calls worst, and the climb is monotonic. A defensive-skill measure would not behave this way. A proxy for shot distance would, and does.
Two repairs that did not work
Comparing within position. The natural fix is to rank centers against centers. It requires trustworthy position labels, and ours were not: the roster data tagged Clingan, Kornet, Neemias Queta, and Robert Williams III as guards. Applying a positional adjustment on top of wrong labels does not remove the bias, it hides it. Anyone attempting this repair should check their labels against a handful of known centers first.
Leaving the defender out of the baseline.The scorer's own average quietly includes the possessions against the defender being judged, which drags every result toward zero. Recomputing the baseline with those possessions excluded is correct and worth doing, but it changes the ordering only slightly. It fixes a real bias that was not the one doing the damage.
A genuine repair needs shot location on every matchup possession, so that a defender is compared against what a typical defender concedes on those shots from those distances. That is a different dataset than the one this leaderboard runs on.
What the data does support
The artifact comes from comparing defenders to each other. Comparing a scorer to himself does not have the problem: if you hold the scorer fixed and ask how he produced against one defender versus his own rate against everyone else, both sides of that comparison are the same player taking broadly the same kinds of shots.
That is the only comparison Edge publishes, and it is why the matchup pages are built per pairing rather than as a ranking. It is a narrower claim. It is also one the data can actually carry.
Even then the sample rules bind hard. Of 21,324 scorer and defender pairings in the tracked data, 390 clear three meetings and 25 matchup possessions. Roughly 98% of pairings do not have enough history to say anything, which is the second reason this section publishes far fewer pages than the row count would allow.
How to check any defender leaderboard in about a minute
- 01
Read the bottom five, not the top five. A plausible top is easy to produce by accident. If the worst-rated defenders are disproportionately centers and rim protectors, the metric is tracking shot location.
- 02
Ask what field-goal percentage each end allowed. If the rating and the percentage climb together, you are reading shot difficulty wearing a defensive label.
- 03
Find the minimum sample. A leaderboard with no stated floor for possessions or games is ranking noise at the extremes, which is exactly where the eye-catching names sit.
- 04
Ask whether the comparison is within a player or across players. Within is defensible. Across players needs shot location, and almost none of them have it.
Method, so you can reproduce it
Source is NBA player-matchup tracking, 24,974 aggregated rows, last refreshed 7 June 2026. Expected points for a defender are the sum, over every scorer he faced, of the possessions between them multiplied by that scorer's points per matchup possession against all defenders. Scorers below 100 total possessions are excluded from the baseline as too thin. Defenders below 150 matchup possessions are excluded from the study, leaving 338. Quintiles are equal-count bands over that set, and the field-goal percentage in each band is the pooled total, not an average of per-player rates.
The possession unit here is a partial matchup possession, only the stretch where the two players were matched, so rates run well below a team offensive rating and are not comparable to one. This analysis was run once, 27 July 2026. It is not recomputed on a schedule, because it is an argument about a snapshot rather than a live figure.
In this section
The comparison that does hold up
Every published pairing, filterable by player and sortable by the gap between a scorer's rate against one defender and his own baseline against everyone else, with the meetings and possessions behind each row.