Arturo from Wages of Wins released his season predictions - Raps looking good

[Craiger]

Yeah. The reason possessions are defined like that is basically because it's more useful to think that way. As you said, if we defined possessions the way you suggest, teams would have different numbers of possessions each game. That would make per-possession efficiency numbers useless until you added in the possessions-per-game stat.

It's usually more revealing, when you're looking at team-level stats, to have a measure for overall efficiency. You can them break that down into its component pieces if you want to drill down and see why the efficiency is what it is. So you see your team has an overall offensive efficiency of, say, 100.0 pts per 100 possession. That's a good measure for comparing how good your offense is overall compared to other teams. But now you want to see what is helping you and what's hurting you, so you check what percentage of possessions end in turnovers, or assists. If you want to know about rebounding, you check to see what percentage of available rebounds your team got, breaking it down into off/def if you want. If you want to know about scoring efficiency, you look at TS%.

Regarding your four points above:

1 & 2: Well yeah, it would be different. But it wouldn't be better, because all you've changed is removing offensive rebounds. You'd have a stat called "possessions" whose only function would be to tell you about rebound differential. We already have that stat, it's called rebound differential. We also already have a stat for scoring efficiency, TS%. And if, for some reason, you really wanted a stat for offensive or defensive efficiency that included turnovers but not rebounding, well, the math is easy enough without redefining the meaning of "possession". Another side-effect of your proposed change would be that any time you wanted to look at per-possession numbers as a measure of efficiency, which is what we use per-possession numbers for most of the time, you'd have to adjust them for rebound rate. It just doesn't make sense.

3 & 4: No, how a possession is defined on the team level wouldn't have any effect on how an individual player is valued. Any box score-based stat that attempts to measure a player's overall efficiency already chooses a weight to assign to offensive and defensive rebounds. Those valuations differ between the different stats, but they all have their reasons (well, except PER, which seems to be pulled straight out of Hollinger's ass).

Bold 1 - I agree, but we could still do this. It would just need to be adjusted.

For example if we use the current equation for a possession using last years numbers.

*All per game league averages

pts per game = 96.3 FGA = 81.4 FTA = 16.9 TO = 14.6 OREB = 11.4

So what we currently get is 96.3/[81.4 + 16.9(0.47) + 14.6 - 11.4] = 1.04 pts per possession

Using the equation I offered we get 96.3/[81.4 + 16.9(0.47) + 14.6 = 0.93 pts per possession

So the league currently averages 104 pts per 100 possessions. Using the equation I offered we get 93 pts per 100 possessions. You can ofcourse do the above and calculate each individual teams average possessions, rebounds per possession, assists per possession etc etc etc. Then compare against the league or another team etc.

Its the same method, with a different equation and therefore different results. So one can still get per possessional efficiency, its just going to be different than what it currently is. (each action per possession will on average be less valuable than how its currently calculated)

I'd argue that accuracy is more revealing than convience. Personally I think measuring the equation by using -OREB is definetely more convient. 1) its always been that way, so continually measuring it that way is convient (especially when comparing against historic numbers). 2) it makes possession equal (in any given game) which makes it easier to compare at a game level.

BUT, if its less accurate than thats not really an excuse. Or atleast shouldn't be. It would be an absolute ton of work to go back and 'correct' old data - yet if we can make that work more accurate, its benificial in the long run

Bold 2 - I'm not saying its necessarily better (I mean I DO think it is, but thinking it is and it being so are two different things). What I really want to know is why does a possession need to be considered over only when the ball changes hands as opposed to when the shooting team gives up control? Or why is that a more reasonable way to value it? It seems rather arbitrary when we consider the rules of the NBA. Its not illogical, after all in most sports we consider a possession over when the ball changes hands. But NBA basketball, with its unique limit on how long one can control the ball (ie. the shot clock) may demand a different view.

Both equations I provided are simple measure of the number of possessions. The say nothing other than 'how do we measure a possession'. But this is also what all box score stats are derived from - its from here they get the estimated value of any actions. So if the average number of possessions per game changes, it will definetely effect the value of all statistics in a game. (as I showed by pts/possession)


Bold 3 - I kind of addressed this already, but it most definetely will effect how each individual players action is valued. Without going into too much detail, how rebounds and missed shots are measured are signficantly different, resulting in much different values (it is one of, if not the, greatest inconsistency between metrics). This would completely alter how PER or WP goes about measuring those particular statistics. I can go into alot more detail into that, but it would have to be extensive and so I don't want to do that without you (or anyone) interested in a more extensive explanation.

[tkfu]

I'd argue that accuracy is more revealing than convience. Personally I think measuring the equation by using -OREB is definetely more convient. 1) its always been that way, so continually measuring it that way is convient (especially when comparing against historic numbers). 2) it makes possession equal (in any given game) which makes it easier to compare at a game level.

BUT, if its less accurate than thats not really an excuse. Or atleast shouldn't be. It would be an absolute ton of work to go back and 'correct' old data - yet if we can make that work more accurate, its benificial in the long run

Neither one is any more or less "accurate" except in some abstract sense of the precise semantics of the word "possession".

I'm not saying its necessarily better (I mean I DO think it is, but thinking it is and it being so are two different things). What I really want to know is why does a possession need to be considered over only when the ball changes hands as opposed to when the shooting team gives up control? Or why is that a more reasonable way to value it? It seems rather arbitrary when we consider the rules of the NBA. Its not illogical, after all in most sports we consider a possession over when the ball changes hands. But NBA basketball, with its unique limit on how long one can control the ball (ie. the shot clock) may demand a different view.

Basketball is different than many other sports. In hockey or soccer, we talk about possession time, not numbers of possessions. That makes sense, because a single team could theoretically control possession of the ball for the entire game. In basketball, though, every time you make a shot you give your opponent the ball. Basketball is more like football in that sense, in that it's about taking turns trying to score. You get the ball, and you're trying to score a bucket before the other team gets the ball. Those are the only two possible outcomes, except at the ends of quarters: either you score (after some number of dribbles, passes, shots, and defensive fouls), or the other team finds a way to take the ball away from you before you score (by stealing your dribble, picking off your pass, rebounding your missed shot, or getting you to commit an offensive foul or violation).

What I'm trying to explain to you is that possessions, as defined by FGA + (0.47)FTA + TO - OREB, is a statistic that only measures FGA + (0.47)FTA + TO - OREB. I've already explained why you use this definition of possession in offensive/defensive efficiency.

I kind of addressed this already, but it most definetely will effect how each individual players action is valued. Without going into too much detail, how rebounds and missed shots are measured are signficantly different, resulting in much different values (it is one of, if not the, greatest inconsistency between metrics). This would completely alter how PER or WP goes about measuring those particular statistics. I can go into alot more detail into that, but it would have to be extensive and so I don't want to do that without you (or anyone) interested in a more extensive explanation.

No, a missed shot is still a missed shot, and a rebound is still a rebound. They are measured in exactly the same way. They are valued differently. But how they are valued has nothing to do with what we're talking about. Both of those stats are individual measures, and the only time the definition FGA + (0.47)FTA + TO - OREB is used is when you are talking about team stats, not individual stats.

[Matt52]

Craiger and tkfu: those are some insightful posts on stats. Thanks.

[minks77]

my eyes are bleeding, guess there's a reason I went to art school.

[Craiger]

Neither one is any more or less "accurate" except in some abstract sense of the precise semantics of the word "possession".

absolutely a 'possession' is abstract, but when it comes to actually applying it to judging it, it requires a precise definition. This is why I used some of the rules of the NBA (24 sec clock, and how fouls are called) as evidence as to how and why we can define a what point a possession ends and starts.

Basketball is different than many other sports. In hockey or soccer, we talk about possession time, not numbers of possessions. That makes sense, because a single team could theoretically control possession of the ball for the entire game. In basketball, though, every time you make a shot you give your opponent the ball. Basketball is more like football in that sense, in that it's about taking turns trying to score. You get the ball, and you're trying to score a bucket before the other team gets the ball. Those are the only two possible outcomes, except at the ends of quarters: either you score (after some number of dribbles, passes, shots, and defensive fouls), or the other team finds a way to take the ball away from you before you score (by stealing your dribble, picking off your pass, rebounding your missed shot, or getting you to commit an offensive foul or violation).

What I'm trying to explain to you is that possessions, as defined by FGA + (0.47)FTA + TO - OREB, is a statistic that only measures FGA + (0.47)FTA + TO - OREB. I've already explained why you use this definition of possession in offensive/defensive efficiency.

I get that FGA + (0.47)FTA + TO - OREB only measures FGA + (0.47)FTA + TO - OREB. But that equation is the core, the starting point, of all non +/- metrics. Every metric uses that number, as you said, to define offensive and defensive efficiency. But it also uses it to 'count' the actions within a possession, and to define what a possession is. My question why? Why that number? Does that actually measure a possession more effectively than FGA + (0.47)FTA + TO? IF we have counted possessions inaccurately, then ALL non +/- statistics are also inaccurate.

Again you may see this as an abstract question, but it has real consequences when we go forward and try to find out how much value any single action has.

No, a missed shot is still a missed shot, and a rebound is still a rebound. They are measured in exactly the same way. They are valued differently. But how they are valued has nothing to do with what we're talking about. Both of those stats are individual measures, and the only time the definition FGA + (0.47)FTA + TO - OREB is used is when you are talking about team stats, not individual stats.

For every action by the offense, there is a counter action by the defense. Thats what these metrics try to quantify. So if a shot is made (FGM), the defense allowed a shot to be made (FGMDef) - the offense gets credit for making the shot and the defense gets credit for allowing them to score. For every FGMissed the defense allowed a shot to be missed (FGMissDef). For every TO, the defense obtained a turnover (TODef). So on and so on. We always have 2 sides to the equation - what the offense did (possession used) and what the defense did (possession gained). Both those sides need to be equal - as defense and offense happen at the same time.

Since FGA = FGM + FGMissed (I'm pretty much going to ignore FTs just to simplify things)

If we are to count a possession used:

Used = FGM + FGMiss + (0.47)FTA + TO - OREB

then to count a possession gained we have:

Gained = FGMDef + (FGMissDef + DREB) + (0.47)FTADef + TODef

So we have 2 sides to the equation.

1)FGM = FGMDef
2)TO = TODef
3)FTA = FTADef

here is the problem spot though

4i)FGMiss = FGMissDef + OREB
4ii)OR FGMiss = Dreb + Oreb
4iii)OR FGMiss = FGMissDef/A + DREB/B + OREB where a+b = 1

where DREB = FGMiss - OREB



Notice something? When we look at 4), we see numerous different ways to count that. We can't count all of them. Different metrics have different versions as to how that should be counted, which in the end is one of the reasons WHY different metrics have different values for the same actions.

--------------------------------------------------------------------------------------------------------------

Ok now I'll try to explain this in a short fashion and use WP and PER as the examples.

*also understand I'm elminating some of the more complex variables here to simplify. This isn't precisesly how the numbers are counted, but this is what they represent without all the considerations to pace, floor time, regression values etc.

So for this example the Raptors are playing Brooklyn. No assists or blocks or FTs happen. X = the 'value' of an action in a possession. That number is equal to avearge pts/possession - to make the example easy we'll say X = 1

Example 1 - Jose makes a shot

PER :

Possession Used - FGM(X) = 1
Possession Gained - FGMDef(X) = 1

The FGM(X) value goes to Jose as he is the one who hit the shot. The FGMDef value goes to the players on the floor for Brooklyn (its divided by 5 and distrubuted to the defenders equally)

WP :

counted exactly the same.

So we have a simple equation with 2 equal sides. That is '1 possession'. Jose made a shot and Brooklyn's defense is responsible. The possession ends and a new possession begins (Brooklyn inbounds the ball).

---------------------------------------------------------------------------------------------------------------

Example 2 - Deron Williams steals the ball from Jose

PER

Possession Used = TO(X) = 1
Possession Gained = TODef(X) = 1

Again rather simple. The TO is creadited to Jose and the steal is credited to Williams

WP

Identical

Again both metrics count 1 possession when the ball changes hands.

---------------------------------------------------------------------------------------------------------------

Example 3 - Jose misses a shot. Lopez rebounds.

PER

Possession Used - (0.7)FGMiss(X) + (0.3)OrebMiss(X) = 1
Possession Gained - (0.7)FGMissDef(X) + (0.3)Dreb(X) = 1

So what we have here is Jose is responsible for the missed shot. The team is responsible for NOT offensively rebounding the ball. Lopez is responsible for rebounding. Brooklyn is responsible for forcing Jose's miss.

*Hollinger's thinking (and I believe this comes from sabermetrics but don't quote me on that) is a rebound is directly tied to a missed shot. After all you can't have a rebound without a missed shot. Averaging O and D rebounds over history we know that approx 30% of all rebounds are offensive (its actually less than this but not important), and 70% of all rebounds are defensive.

So since there is only 1 possession, and a rebound is a result of a missed shot - both the rebound and the missed shot are a % of the same possession. Since 30% of missed shots are offensively rebounded (and would therefore NOT end the possession but rather continue it), 70% of the possession goes to the person who missed the shot. Therefore the remaining 30% goes to the rebounder. Hence the (0.7)FGMiss and (0.3)Dreb. I hope that makes sense as it is VERY important to how and why these metrics count differently.

WP

Possession Used - FGMiss(X) = 1
Possession Gained - DReb (X) = 1

Jose is responsible for the missed shot, Lopez is responsible for the rebound.

We can see that WP is much more simple in how it counts this possession. Jose misses the shot so he is responsible for the lost possession. Lopez grabbed the rebound so he is creditted for the possession.

But notice something. When Jose hit the shot in example 1, the defense were all responsible for allowing him to hit that shot. Here we have a missed shot, but the defense is not responsible for that missed shot. Rather all the responsibility is given to Lopez

--------------------------------------------------------------------------------------------------------------


See the significant difference between how each metric 'counts' a possession? The exact same play (example 3) has 2 very different ways to be counted and as a result all the players on the floor receive different values.

PER thinks a rebound is directly tied to a missed shot. The missed shot and the rebound are both parts of the same action.
WP thinks the rebound is iresult of the missed shot. But in order to 'count' that rebound, it eliminates the FGMissDef part of the equation. You can't have BOTH DREB and FGMissDef as non-fractional numbers otherwise:

Possession Used : FGMiss (X) = 1
Possession Gained : FGMissDef(X) + DReb(X) = 2

Since 2 does not equal one, that equation is mathematically impossible. Therefore, under WP one of FGMissDef or DReb have to = 0.


Now why is this significant? PERs conclusion that a rebound and a missed shot are both parts of the same possession is not necessarily wrong. BUT deciding a (0.7)FGMiss and (0.3) Dreb is rather arbitraty. It is most definetely based in reason, but there is nothing that says THIS IS TRUE. Its a guess. Also it says different rebounds have different values. An offensive rebound > defensive rebound. But both are the exact same action with the exact same result, so how or why should an OREB be worth more than a DREB? WP on the other hand doesn't 'guess'. But in order for their possession gained equation to work, one of DREB or FGMissDef has to be = 0. So either the defense has value or the rebound has value, but they both can't have value at the same time.

*this is WHY rebounds are considered 'overvalued' under WP.

This all stems from the equation of FGA + (0.47)FTA + TO - OREB. It directly ties ALL missed shots and rebounds together as part of 1 possession. But it results in a very complex and diverse way to 'count' those rebounds, missed shots and the defense.

--------------------------------------------------------------------------------------------------------------

So when we look at what I'm thinking

Offensive Possessions = FGM + FGMissed + (0.47)FTA + TO
Defensive Possessions = FGMdef + FGMissDef + (0.47)FTADef + TODef

1)FGM = FGMdef
2)TO = TODef
3)FTA = FTAdef
4)FGMiss = FGMissDef

no need to 'guess' at what 4) should be. Any missed shot is entirely the responsibility of the shooter, and on defense its entirely the responsibility of the defense. Rebounds don't need to be directly linked to the missed shot. They are independant actions, much like assists or blocks. And theory says this could be true, as for every missed shot there are 10 possible rebounders, each having to attempt to get that rebound. (ie. only one person can miss a shot, but 10 different people can rebound that shot)

Now thats not to say my idea is complete. Clearly as it stands rebounds don't have a value. But, this is just a stepping stone idea. Much like assists and blocks are added in addition to a possession used or gained (as you can see in the equation for a possession they don't even exist), rebounds could be aswell.



I hope to God (or whoever) that makes sense and I explained that reasonably well (didn't miss too much or make too many errors as a full breakdown would take hours and hours. I didn't even touch what happens when a shot is missed and is offensively rebounded). But I'll admit I'm no teacher.

[Fully]

I'm generally a supporter of the advanced metrics movement in basketball, but I feel like it should be part of a bigger formula when determining a players worth. Extracting a conclusion from strictly advanced stats is as silly as extracting one from all the raw per game averages that this community seems to stick their nose up at.

All you have to do is take a look at the results and common sense would tell you that something is amiss here. Kawhi Leonard, as valuable and efficient as he is, is not a top five player. Brandon Wright and Matt Barnes are not top 20. Brandon Rush, Kosta Koufos and Danny Green are not top 30... That's not really up for debate. I will be honest - when I see findings like these, I feel like it actually hurts the credibility of advanced metrics instead of helping it.

[Brandon]

WoW would consider the Raps the favourites to take it all if they'd change their team to include only big forwards who rebound and never shoot unless it's a dunk. Reggie Evans, for example. 15 guys like that and you've got your WoW championship.

[Craiger]

WoW would consider the Raps the favourites to take it all if they'd change their team to include only big forwards who rebound and never shoot unless it's a dunk. Reggie Evans, for example. 15 guys like that and you've got your WoW championship.

not exactly.

If everyone never shot unless it was a dunk, there would either be alot of dunks (high% shots) or alot of turnovers. If all those guys were able to dunk on almost every possession, then yes they would be very good. But any metric would say that to. If they didn't shoot unless it was a dunk, and there were very few dunks available, then they'd turn the ball over alot (24 sec shot clock). Which would then mean all those players would look very bad.

As there is a limited number of rebounds in a game (limited to the number of shots missed), rebounds would either be distributed equally or unequally. So either everyone grabs close to an equal number (and brings down everyone average) or people grab rebounds unequally (and bring down only a few players averages). And ofcourse if everything was a dunk - there would be very few offensive rebounds.

WP also has the positional adjustment, so as some of those guys would have to PG etc, so some of them could look 'worse' if they weren't doing what a PG 'should' do.

Its only because the 'low usage high rebounding' guys are distributed amongst many teams that they look 'better than they should'. If they were all put together on one team, either all of them would look worse, or some of them would be terrible.

edit : I should add that WP at face value may say as a prediction that said team would be a contender going into the season, but as the season progressed it would say something significantly different. Thats one of the 'problems' with WP. Its not very predictive, but rather accurate (atleast in relations to how they see 'wins') when we look at what did happen. If things stay relatively constant from year to year (which they do) then we can use that commonality of 'what did happen' to help predict what will happen. However, any significant change (such as having 15 roster spots filled by big men) will lead to very significant different result.

[slaw]

All you have to do is take a look at the results and common sense would tell you that something is amiss here. Kawhi Leonard, as valuable and efficient as he is, is not a top five player. Brandon Wright and Matt Barnes are not top 20. Brandon Rush, Kosta Koufos and Danny Green are not top 30... That's not really up for debate. I will be honest - when I see findings like these, I feel like it actually hurts the credibility of advanced metrics instead of helping it.

A comparison to baseball is good. This year, as a result of the defensive shifts the Jays were employing, Brett Lawrie basically "broke" the defensive side of the WAR equation. Instead of its adherents continuing to argue that the metric was perfect (as the WoW guys do) and shouting down people who objected as ignorant (as the WoW guys do) everyone just acknowledged that WAR wasn't perfect (no, Lawrie isn't as good as Babe Ruth) and moved on to try and fix it.

The problem with the WoW guys isn't so much their metric as how they use it. I read a great line about Dave Berri once and it was something like: when all you have is a hammer, you think everything a nail. Berri's hammer is regression analysis and he goes about beating everything to death with it until it turns into something vaguely nail-like".

All I say about statistical analysis in sport is that when you are trying to measure incredibly complicated things with regressions, a little humility and caution are in order. The WoW guys too often display neither.

[Hugmenot]

not exactly.

If everyone never shot unless it was a dunk, there would either be alot of dunks (high% shots) or alot of turnovers. If all those guys were able to dunk on almost every possession, then yes they would be very good. But any metric would say that to. If they didn't shoot unless it was a dunk, and there were very few dunks available, then they'd turn the ball over alot (24 sec shot clock). Which would then mean all those players would look very bad.

I believe Brandon's comment was more than likely made in a very sarcastic tone.

Your second paragraph highlights a fatal flaw with Wages of Wins when applied to the player level. A player is not in full control of whether to shoot and from where, far from it. WoW seriously overvalue players who can't shoot nor drive, even more so if they are good offensive rebounders.

About the change of possession formula. HoopsData uses 0.44 instead of 0.47. I've also seen discussions that uses other values. I am not a proponent of such measures and thus I decided to use concepts which are typical for risk-adjusted morbidity and mortality models when I started doing research on my own last year. It gives me something to do while planning for my retirement.

[Craiger]

I believe Brandon's comment was more than likely made in a very sarcastic tone.

Your second paragraph highlights a fatal flaw with Wages of Wins when applied to the player level. A player is not in full control of whether to shoot and from where, far from it. WoW seriously overvalue players who can't shoot nor drive, even more so if they are good offensive rebounders.

About the change of possession formula. HoopsData uses 0.44 instead of 0.47. I've also seen discussions that uses other values. I am not a proponent of such measures and thus I decided to use concepts which are typical for risk-adjusted morbidity and mortality models when I started doing research on my own last year. It gives me something to do while planning for my retirement.

Nice pick up. WP uses 0.45. I'll admit to only quickly checkign that number. But its perhaps the least signifcant number that could be wrong in that entire equation

And I agree in part.

Average fg% is 46%, and to be a net positive when a player shoots they need to shoot 50% under WP. That would indicate shooting % is valued to highly. But whats the magnitude of that? Just how much is it overvalued? Enough to make the statistic to 'wrong' to have meaning?

We can take that further, how much of a negative impact does 'not shooting' have on a team? Why is a player not shooting? Why is another player shooting more? Is a player shooting less than average not having a negative impact on the team?


So while arguing WP 'overvalues' shooting may not be wrong, its also not necessarily right. There in lies the rub.

I just get the impression that the most common reason why someone doesn't like stats (or a stat) is pure confirmation bias. "This # doesn't match up with what I see or what I think I see". But that doesn't mean its wrong or come to any understanding of how wrong.

[Puffer]

It is amazing to me how many posts ago I stopped being able to understand this thread. There are some talented people posting on this board. Of course, being talented doesn't stop them from saying some really stupid shit sometimes.

[Kuh]

An amusing note: in the first three games this season:
- Andrea shot 10/32
- Amir+Ed+AGray shot 18/35 in twice the minutes

Of course, this is a really small data set. And it makes the point of the WoW guys very well: sitting Andrea for the 'guys who can't shoot' would on these three days have given us roughly the same number of baskets and points (getting to the line is also a wash: AB had 7 made free throws, AJ+ED+AG had 12 in twice the minutes). The 'can't create their own shot' brigade would have done this from 17 rather than 32 field goal attempts ...

... in this tiny sample, it looks like the biggest complaint against Amir and Ed is not that they don't score enough points: they match Andrea. The biggest complaint is that they don't miss enough shots.