The GOAT Ladder: Defensive Greatness

Defense – The Final Frontier

It has heretofore been impossible to discuss individual defensive performance with any type of common language. Unfortunately, that leaves us with no solid means to debate the defensive merits of the majority of players in NBA history. All previous attempts to analyze defensive value have relied on defensive rebounds, steals, and blocks. In their absence, some models have used rebounds and assists as proxies. Statistics built upon this foundation include Dean Oliver’s DRating, bball-reference’s Defensive Win Shares, and Myers’ D-BPM. Some all-in-one metrics have also used team Defensive Rating as a stand-in for a player’s defensive value. DRating, Wins Produced, and logically even Plus-Minus models (in so far as points allowed is half of a team’s point margin) all use team defensive strength in this way.

The result is a decided favoritism toward big men (who gather rebounds) and players who are on good defensive teams. Many have noticed the result, but have not been able to redress the deficit without the use of superior data. Analyzing defense in this way also substantially misrepresents the merits of perimeter defenders. Such methods reward players with a lot of steals (even if they are poor defenders who gamble a lot) and punish players who don’t get many steals (even if they spend a lot of their time guarding the opponents’ best player.

What Have You Tried?

I have attempted to rectify this state of affairs by developing Matchup-Based Defense. Since the 2013-14 season, the NBA has used tracking technology to record who each player is guarding on a regular basis. This is far and away the most substantial basis for any defensive analysis, and allows us a unique insight into the modern game. The results of this analysis are on the homepage, and have been the basis of several posts and a series of visualizations. If you wish to see a more detailed breakdown of a player’s performance according to Matchup-Based Defense, there is also a Google Sheet that allows the user to select player and season and get a complete analysis.

As valuable as the matchup data is, however, it account for only a tiny pocket of NBA history. You may be convinced, as I am, that the Matchup-Based Defense model tells us quite a lot about individual defense in the last few seasons, while still wondering how much we really know about the rest of NBA history. What about everybody who played before 2013-14? There is no tracking data from their games, so how might we determine who played good defense and who played bad defense?

Where Does That Leave Us?

In the big picture, we have an incomplete picture of most players in NBA history. Despite having a good idea of how valuable a player is on offense, we have little clue how much a player helped or hurt his team on defense for about 2/3 of NBA history. We thus have no way to know who is the greatest player in any season prior to 1996-97. As a result, we cannot measure the greatness of any player who played in any season prior to that time.

That time is over now.

Using a process that is easy to understand for most basketball fans, I have been able to evaluate individual defensive performance by every player in every season since 1952-53. Here’s how:

We begin by making the best guess possible about which type of opponents a player guarded. For most of NBA history prior to about 2016, players of similar sizes defended each other. In truth, this fact is basis for the common language we use to talk about “positions”. Tracking data reveals that this is no longer the case in the pace and space era. About 90% of the NBA’s history fits this description, however. Centers guarded centers, forwards defended a similar-sized forward, and guards defended a similar-sized guard.

“Who’s Your Man?”

In order to estimate which opposing players a defender covered, I started with position tags from basketball-reference. I then transformed the labels by assigning “multi-position” players to the traditional position which most closely aligns with his defensive role. A big player who is a jump shooter may be labeled as something like “PF-SF”, “F”, or even “SF-C”. These designations are accounting for the player’s offensive style as well as his defensive role, though. The most important feature for recovering the player’s defensive assignment is recognizing that he is physically a “big man”. Although he may function like a forward on offense, the player is likely to defend opposing big men. His size (and, we may infer, probable lack of footspeed and lateral quickness) make it a necessity.

Similarly, a player may be labeled as “PG-SG” or “G” or “SG-PG” if he shares ballhandling duties with his backcourt-mate. A bigger ballhandler may even be labeled as “SF-PG” or something zany like that. What these labels tell us is that the player is probably not the smallest guard in his team’s lineup. Though he may function as “point guard” on offense, he is likely bigger than his counterpart guards. Thus, he is likely to defend a bigger opposing guard.

Using this set of assumptions, I reduce every player’s tag to a single position which best represents his defensive role. We then evaluate the performance of all players on the opposing team who have the same position label. Then we do the same thing for each game the defender played in that season.

What Is Good Defense?

How can you tell what effect defender had on his opponent? If Michael Jordan puts up 22 Points Created in a 1991 game against the Detroit Pistons, who have Joe Dumars Vinnie Johnson, and John Long at SG, does that mean that Detroit’s shooting guards played good defense or bad defense? Superstars score on everybody, good defender or bad. Look at the barrage of strong defensive forwards Denver threw at LeBron James during the Western Conference Finals – then look at the results. You might well think nobody was guarding LeBron at all, with how dominant he was on offense!

On the other hand, teams count on their superstar to carry the burden for their offense. The star player needs to create a lot of offensive value in order for his team to be successful. Is it fair to compare him to average, and then punish the defender(s) who let him score more points than average? Surely not. The opposing team knows that the star player is his team’s focal point on offense. They will likely deploy the best possible defender to cover him. Evaluating the star’s performance against league average will lead us to the incorrect conclusion that the defenders guarding him played badly.

On the other end, consider the defender who is facing off against Frank Ntilikina. Ntilikina may have a measly 5 Points Created in the game under consideration, falling well below average for his context. Does that mean his defenders (let’s say it’s the Washington Wizards point guards) did a good job? Not really. We don’t anticipate Frank Ntilikina making much offensive impact on any game. Holding him to a low output doesn’t tell us much about the players defending him.

Calculating Defensive Impact

We measure defensive impact by comparing opponent performance against that offensive player’s average performance. Playing good defense against Michael Jordan means holding him below his high standard (even though he will likely still have an above league-average performance). Playing good defense against Frank Ntilikina means holding him below his own meager standard. Allowing Frank Ntilikina to put up a league average performance is actually bad defense.

Thus, we find a position unit’s defensive performance by subtracting observed performance minus expected performance for the other team’s corresponding position unit. In the first example above, we compare the total performance of Chicago SGs against their expected performance. The expected values take into account:

• Minutes played by each player
• The pace of that game (and thus, how many possessions each player was on the court for)
• Each player’s average performance per possession

To be more specific, we calculate each offensive player’s expected statistics separately, then aggregate them with teammates at the same position. Then, we divide the weighted total by total possessions played by that position unit in the game.

Subtracting the expected output from the observed output tells us whether or not the offensive players performed better or worse than they normally do against these particular defenders. The difference is the raw defensive impact of the position group (Detroit SGs, in our example). Each defender in the position group receives credit or blame for the opposing unit’s performance based on the following attribution method:

Dealing With How Much of the Game Each Defender Played

For the example above, we find both the total output of Chicago SGs in their game against Detroit, and the expected output of Chicago SGs based on the per-possession performance of Chicago’s SGs, the number of possessions played by each SG, and the pace of the game. We subtract the expected performance for Chicago SGs from their actual performance to determine how well or how poorly they were defended. The result, whether positive or negative, we divide among Detroit’s SGs based on how much of the game they played.

The amount of the game each player played is equal to (MP / (Team MP / 5)), or MP/48 where box score data is incomplete. This accounting workaround is necessary to deal with cases where too many players on the same team have the same defensive position. In such cases, Detroit SGs might well have played a total of 80 or 90 minutes. Since this happens as a result of either players being mislabeled or unusual short-term rotations (due to injury or suspension, for example), dividing a player’s minutes by the true maximum amount of minutes available for his spot is a failsafe to prevent a defender getting more or less credit (or more/less blame) than he deserves for a game simply because of a bookkeeping irregularity.

Comparing Defensive Value to League Average

On average, defenders hold offensive players to … exactly their normal level of performance. There are good defenders, bad defenders, and everything in between, but the average level of defense is when the opposing team performs exactly as expected. As such, league average defense would produce a differential of zero over the course of the season. Better defenders will produce a differential far below zero (opponents perform worse against that defender than they normally do), and bad defenders will have differentials far above zero (opponents play much better against that defender than they do the rest of time).

Now we have the same problem we had when comparing players’ offensive performance against league average: the players at the top and bottom of the totem pole in high-scoring eras appear to be farther away from average than the players at the top and bottom of the league in low-scoring eras. Because of what we know about variance (and the amount of rain in Seattle and Phoenix), we know that this is deceiving. The truth is that players will tend to be farther away from average when more points are available.

Fortunately, the same process we’ve used for scaling offensive statistics will work just as well for defensive statistics. Just like before, we calculate how far away the average player is from league average. For some seasons that might be 20 points, or in some seasons it might be 100 points. Then we divide the player’s differential by this value. The result tells us how many times farther away he is from the middle of the league than an average player is. A really good defender might be 2 or 3 times farther away from league average (below league average) than his peers. A terrible defender might be 2-3 times farther away from league average than everybody else, except that he will be that far above league average. He improves his opponents’ performance by way, way more than the average defender would.

Negative Numbers are Good, Actually

For each defensive statistic we now have a standardized score. We have scores for my metric (Opponent Points Created) or box score stats (Opponent PTS, FGM, FGA, FTM, etc.). We can use this score to compare the defensive impact of players who played under very different circumstances. The important thing to recognize is that when we’re looking at numbers that measure impact on opponent’s performance, it’s better to have a negative score than a positive score.

D_Score, my season-level defensive metric, expresses the impact a defender has on his opponents’ Points Created. The best defenders have a low value (actual value – expected value = a number less than zero), while the worst defenders have a very high value (actual – expected gives a result that is greater than zero. To give you an idea of what a normal value is for D_Score, here is the distribution across NBA history:

The same holds true for several other scores: OPP_PTS Score, OPP_FGM Score, OPP_FGA Score, OPP_FTA Score, and OPP_OREB Score. For these values, negative numbers are good. These values appear in bar chart form on the Defensive Opponent Impact dashboards here and here. Since the scores represent the defender’s effect on his opponent’s performance, more effective defenders have a higher expected value than actual value. The opponents that the player guarded had less of whatever statistic we’re investigating than he normally would have. Having a negative OPP_PTS Score means that opposing players scored less against this defender than they would normally score.

But Then Again, Positive Numbers are Good Too

There are cases where positive numbers are good on defense. OPP_MISSES Score, OPP_TOV Score, the defender’s own BLK Score, STL Score, and DREB Score are all higher for better defenders. This second group of scores are recorded on the “Direct Defensive Impact” page of the Historical Dashboard found here. They represent positive contributions that we can directly connect with a defender. If a player has more of these events than expected, that’s a good thing.

In essence, I’m claiming that there are two ways to detect a player’s defensive performance in the statistical record: 1) the plays that he makes (forcing a miss or turnover, or blocking a shot), and 2) preventing his man from making plays.  

The most significant among this group of scores, in my mind, is OPP_MISSES Score. The Opponent Impact Scores (particularly OPP_PTS and OPP_FGM) can be impacted by fluctuations in opponent volume. If a defender is guarding a high-volume shooter, he may well score a lot of points, but do so inefficiently. The defender may force him into difficult shots and generate a lot of misses, even though his OPP_PTS Score will look bad. OPP_MISSES Score helps to give us a more complete picture in such cases. Indeed, we will find that when defender has to guard mostly high-volume shooters, he will produce a very high OPP_MISSES Score in spite of having a pedestrian OPP_PTS Score. This ensures that we recognize good defenders, even when they’re guarding a player who shoots a lot.