APBRmetrics

For a long time I always believed that clutch play has no correlation to future clutch performance due to the random nature of clutch games.


I decided to run correlations between clutch play (defined as performance when the game is +/- 5 points and 5 minutes left in the game) between pre all-star break and post all-break for the seasons between 2005-2014.

R values:
W/L% 0.374
MIN -0.049
OffRtg 0.274
DefRtg 0.200
NetRtg 0.293
AST% 0.280
AST/TO 0.146
OREB% 0.136
DREB% 0.050
REB% 0.151
TO Ratio 0.135
eFG% 0.178
TS% 0.210
PACE 0.219

Team performance in the clutch actually correlates well with future clutch performance. It even performs better than stats like Rebounding and pace which typically have high correlation values.

What's interesting to me is that the correlation for W/L% in clutch games has a year to year correlation of .380. 108 out of the 150 (72%) of the teams who finished a .500 or better clutch W/L% matched that feat in year N+1. 64% of the below .500 clutch W/L% teams finished below .500 in clutch W/L% the next season. This might explain why LeBron's teams and Dirk's teams always finish well in clutch situations. It's not a fluke, but a skill.

I'm in agreeance with this.

Seems to agree with what we already know about teams performing better when they're down.

Here's a fun DeanO article that brushes on the subject, though indirectly: http://www.rawbw.com/~deano/articles/aa052097.htm

What are we really measuring here? I'm sure that good teams will play better in any given five-minute stretch of the game, and that will correlate to some degree before/after All-Star break, from one year to the next, etc.

But what "clutch" truly means is an ability to outperform/underperform overall ability in important situations. That's the tendency/skill we'd want to test for, not simply playing well in important situations.

I also would be careful with the win-loss in "clutch games" because I believe a game where a team leads by five at any point in the last two minutes is treated the same as a tie game. But that's certainly not a 50-50 scenario. A win is much more likely than a loss at that point, even though it's not certain.

Kevin Pelton wrote:What are we really measuring here? I'm sure that good teams will play better in any given five-minute stretch of the game, and that will correlate to some degree before/after All-Star break, from one year to the next, etc.

From what I read, it was conventional wisdom that close games are 50-50 coin flips and that teams that win close games are unlikely to be that good in close games in the future.

For example here is an excerpt from B-R:

Although APBRmetrics teaches us that all teams' records in close games regress toward .500,

http://www.basketball-reference.com/blog/?p=9018

colts18 wrote: From what I read, it was conventional wisdom that close games are 50-50 coin flips and that teams that win close games are unlikely to be that good in close games in the future.

"regress toward 0.500" does not mean that something becomes a "coin flip", maybe like each event in basketball as a "coin flip" with a loaded coin, but not with a fair coin. A better team is still expected to win close games at a higher than 0.500 rate; the expected rate of winning such games just goes down towards 0.500 for a better team and goes up towards 0.500 for a worse team, because in such scenario "luck" plays essentially a bigger role. Nonetheless, the better team is expected to win more often than the worse team.

It also makes not much sense to compare win% here, because as Kevin Pelton pointed out, the probability of winning a game changes with the time remaining on the clock as well as the existing scoring margin. A team leading a "close game" by 5 with 10 seconds left is more likely to win that game than the team trailing by 5 with 10 seconds left. What might be a starting point for whether (or how much) skill is involved is the respective scoring margin created under such conditions. How does that relate to the team average scoring margin created?
Also, we are dealing with a huge sample size issue, because there are clearly less minutes under the condition "close game" than all other minutes. Making it really tough to actually have signficant results for such analysis. Doing such analysis just for teams alone, might also give a misleading result due to the fact that teams may have played different lineups in such minutes (maybe due to injury, due to players being fouled out, substitution patterns, etc.), and we know that the playing level depends on the used players as well as the opponents. So, overall, such analysis would be best done by using the respective lineup data in such situations. Then we can draw conclusions for each player.

In a big enough sample the results for such "close game situations" for each player should regress towards the typical performance level of each player, which still means that with a lineup with better players should win at a higher rate than a lineup with worse players. Which doesn't mean those players are "clutch", but rather simply better players.

colts18 wrote:From what I read, it was conventional wisdom that close games are 50-50 coin flips and that teams that win close games are unlikely to be that good in close games in the future.

I'm not sure what conventional wisdom is at this point, but I wrote about this a few years ago:

I wrote:There are two extreme schools of thought on close games--those that believe they are primarily decided by luck and those that feel they are primarily decided by teams and demonstrate their true ability. Neither position is supported by the data.

Instead, what the results tend to show is that the difference between good teams and bad teams is mitigated in close games. Look at the best-fit regression line on the chart. The slope is nowhere near 1, and the difference between the expected record in close games for the very best teams (about .600) and the very worst teams (about .400) is much smaller than the difference between them in games that are not decided down the stretch.

Which is, I think, consistent with what you show, the difference being the way StatsCube does win-loss record in clutch games makes these margins look larger than than a method based on final scores. But here's the important part:

I wrote:A team that outperforms its expected record in close games is not necessarily more likely to do so again the following season than one that underperformed.

There's the rub. I suspect there is a real "clutch ability" for certain teams/players, but only in the most extreme cases (the Love/Rubio/Adelman Timberwolves being an example) are the sample sizes large enough for these differences to reveal themselves.

I think if you’re a 90 percent kicker, the best you can do under the most important situation is be a 90 percent kicker. I think if you’re trying to stretch any more than that … that’s not realistic. ...
-- Steven Hauschka (kicker for the Seahawks)

I'm not a huge believer in "magic clutch performance." It seems much more likely that teams and players pace their effort and are operating at closer to 100% when the game is on the line than when they have a large lead.

From what I read, it was conventional wisdom that close games are 50-50 coin flips and that teams that win close games are unlikely to be that good in close games in the future.

If the only information that you're using is the final score, then, if the final score is close, that suggests that the teams are close in quality, but if you're looking at other stuff, the same game might look like a dominating performance.