Page 1 of 1

eWins and other accrediting systems (MikeG, 2010)

Posted: Fri Apr 15, 2011 9:40 am
by Crow
Page 1

Author Message Mike G



Joined: 14 Jan 2005
Posts: 3615
Location: Hendersonville, NC
Posted: Mon Sep 24, 2007 6:43 am Post subject: eWins and other Individual Wins accrediting systems

(This thread is a spinoff from another.) I'm feeling prompted (again) to ask a rather open-ended question. In placing individual (player) accountability for (team) wins, should individual wins add up to... 1)... actual team wins? 2)... team statistical performance, i.e., pythagorean expected wins? 3)... neither? I think WS does #1, and PW shoots for #2. When I did eW, I didn't know if either was the better target. If you are considering individual wins (IW, to be generic) to be a stat that players accumulate, then the best teams -- which may win 2-3 times as many games as the worst teams -- might have 2-3 times as many IW distributed among their rosters. But an elite team may only score 15-20% more points (as differential) than a cellar-dwellar. All the things that go into point diff -- TO, Reb, Ast, Stl, Blk... -- are generally also within a similar range. If IW are a quantity of something real, and the bottom-line team quantities -- points scored -- are within 20% of one another, then is it practical to suppose players on a weak team are really only generating half (or less) of something than those on better teams? The pat answer is, "They're generating only a fraction of the wins". Sure, but the stats they generate indicate that they're making their team more competitive, more often within striking distance, with a good rally or defensive outburst. In fact, we might define all player contributions as being stratified into levels that (1) make their team less terrible, (2) make them competitive, (3) propel them to wins. I propose that any system forcing player IW's to sum to team wins (or PythWins) requires artificial support to do so. Players are deemed to be from excellent to terrible on defense, basically. Yet the same players don't carry over their excellence (or lack thereof) from good teams to bad teams. This doesn't seem intuitively believable to me. The eWin system doesn't suppose that a players' defense becomes many times better or worse when he changes teams. They also don't add up to team wins. Team eW still add up hierarchically so that the best teams are best. But they aren't 2-3 times as good as the bottom teams; only some 50-60% better. Everyone has a different intuitive sense of 'good' and 'bad'. At one time, a 40% FG rate was 'good'. At this point, I won't even get into 'individual losses'._________________` 36% of all statistics are wrongLast edited by Mike G on Mon Sep 24, 2007 10:02 am; edited 1 time in total
Back to top

deepak



Joined: 26 Apr 2006
Posts: 665
Posted: Mon Sep 24, 2007 10:01 am Post subject:

Quote:
If you are considering individual wins (IW, to be generic) to be a stat that players accumulate, then the best teams -- which may win 3-4 times as many games as the worst teams -- might have 3-4 times as many IW distributed among their rosters. But an elite team may only score 15-20% more points (as differential) than a cellar-dwellar. All the things that go into point diff -- TO, Reb, Ast, Stl, Blk... -- are generally also within a similar range. If IW are a quantity of something real, and the bottom-line team quantities -- points scored -- are within 20% of one another, then is it practical to suppose players on a weak team are really only generating half (or less) of something than those on better teams?
Aren't you assuming that IW can be formulated as a linear combination of box score stats? As I see it, wins are only indirectly determined by such stats. Something like: stats -> point diff, possessions -> estimate for wins Edit: Hmm, nevermind. I misunderstood your point.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3615
Location: Hendersonville, NC
Posted: Mon Sep 24, 2007 10:19 am Post subject:

Yes, if I understand what you're saying. eWins takes boxscore stats, scales them to Team/Opponent productions, and then predicts Wins non-proportionally. Expected team wins (for an 82G season) are: xW = eW*2 - 41 So a 61-win team typically totals 51 eW. A 21-win team totals 31 eW. The good team isn't 3X as good as the bad one, but only 1.65 times as good. Here, 'good' is effectively defined as 'total talent above replacement level'._________________` 36% of all statistics are wrong
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Mon Sep 24, 2007 10:24 am Post subject:

An Individual Win metric probably should compare to league average or replacement level on stats or to average or replacement level player. I'd be wary of one that didnt and didnt fully. It is the difference from these baselines that affects winning and losing not the total stat level or total player level. EWins compares to replacement player. From your formula in the other thread it looks like a T=15+ for 2000+ minutes starts to contribute eWins. Not sure if T=15 is at or above/below mean or how closely eWins tracks with PER but does it in fact boil down to saying it takes a statistically above average player to contribute individually to wins? In reality players are the sum of their individual stats and team effects above and beyond those individual stats. Dan's overall +/- has both (and from offense and defense) with individual generally the far heavier weight though it probably varies by type of role player. The pure +/- tries to capture the uncounted team effects. This metric appears comprehensive in scope. Berri's Wins Produced uses league average performance on shooting. That isn't a terrible choice in itself but maybe the mistake is the mix he sets it in, one that doesnt not do the same for rebounds, etc. at least for Wins Produced. His PAWS/min uses replacement player. PAWS/min is more palatable to me as positional average performance is a main design feature and the shooting formula decision at least hits all the players at the position the same but it might be more acceptable to many here to re-run Berri's formula with shooting break-even at replacement player shooting level or some more neutral or statistically driven level. Anybody done or interested in doing this? Berri's metric does incorporate the impact of overall team defense on scoreboard but in a crude equal way (something PER is lacking, but just as crude as player W-L off of OR and DR ratings). Is it missing the offensive team effects of players or is that getting mixed in with the "team adjustment"?Last edited by Mountain on Mon Sep 24, 2007 10:39 am; edited 3 times in total
Back to top

deepak



Joined: 26 Apr 2006
Posts: 665
Posted: Mon Sep 24, 2007 10:32 am Post subject:

I think this really depends on what we're trying to describe: (a) Do we want to measure how much credit a player should get for his team's success? If so, then it makes sense for the IWs to add up to team wins (or expected team wins). (b) Or do we want to measure player performance, irrespective of his team's success? If so, then the eWins system makes more sense to me. More practically, I think we'd like to know how much a player would contribute to the success of a future team. That is, if I have a new roster I put together, how successful can I expect this team to be, and how much would each player contribute? For this purpose, which type of metric would be more useful -- (a) or (b)?
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Mon Sep 24, 2007 10:49 am Post subject:

The 2:1 ratio of wins / statistical performance difference you observe with EWins Mike, I have observed with team minutes weighted tendex and team PER too so this appears to a rough universal rather than system specific. And caused by crediting possession affecting actions as well as scoring actions. I think Nick S spent a good of time talking about accounting and touches on this.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3615
Location: Hendersonville, NC
Posted: Mon Sep 24, 2007 10:56 am Post subject:

Mountain: Of course average players contribute to wins. Replacement T was figured at 11.7 last year; over 400 players met this level, 13-14 per team. All contribute a net positive. The 30 top 'sub-replacement' types (in minutes played) averaged 37%TS. 30 guys around 11.7 (15+ and 15-) averaged .433. Last I checked, a PER of 15 was roughly equal to a T of 26. This varies from team to team, as T rates actually correspond to team strength. Deepak: eWins ranks players on a given team; thus gives due credit for that team's success. It also gives a Kevin Garnett (great player/terrible team) with more 'credit for success' than a Jason Terry (good player/great team) -- since KG takes an otherwise-NBDL lineup and makes them competitive -- unlike some other IW systems. If you are building a team and adding last year's player eW rates (properly allotting minutes/positions), you will do very well in predicting this year's wins. PW and WS, as far as I know, make no claim along these lines; much of their credits are tied up in 'team defensive' schemes. Remember, eWins is very cognizant of team success. The factors of Tm/Opp Pts and Reb are what distinguish it from others. But the scaling is much less drastic than what is found in W-L records._________________` 36% of all statistics are wrongLast edited by Mike G on Mon Sep 24, 2007 11:10 am; edited 1 time in total
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Mon Sep 24, 2007 11:07 am Post subject:

Mike, my bad, I was thinking of wins against an average team level, 41, (as in some metrics I have worked on conceptually) rather than wins period. Certainly players below average will contribute to wins period. Different scale. So a TS+26 playing starter minutes would contribute somewhere in the range of 6-8 wins. that makes sense. I like the innovative features of EWins, hence supported scoring it against other more widely publicized measures.Tm/Opp Pts does incorporate shot defense (with everything that affects opponent points), though the team defensive result is assigned to all equally, correct (as in some other systems)? Would you have any interest in modifying to base it off of say a 50% weight of team opponent scoring and 50% estimated counterpart scoring? Giving EWins another distinction, that perhaps puts it closer to what is happening by recognizing at least some- call it half to start- of shot defense is local?Last edited by Mountain on Mon Sep 24, 2007 2:19 pm; edited 1 time in total
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3615
Location: Hendersonville, NC
Posted: Mon Sep 24, 2007 12:13 pm Post subject:

Mtn, I greatly appreciate the shout-out, as I don't really think eWins is 'well known', unless you are a thorough reader of these forums. Also thanks for reminding of Nick S's work a few weeks back; I was heavily curious where he would take that. Dan D (Statman) is another to watch. But I'm very skeptical of 'counterpart' effects. The Spurs don't match up with others' positions, and they're the flagship franchise of the NBA. Coaching quirks are unpredictable. Players don't have positions, on O and on D, in my view, and the best players will always be just Players. Fixing things at the 90% level is a lot closer to 100 than to 50. That's how I think of team effects: You can go to many times more work for the next few % of accuracy. You still win by outscoring your opponent, and the points-ratio is what determines wins._________________` 36% of all statistics are wrong
Back to top

NickS



Joined: 30 Dec 2004
Posts: 384
Posted: Mon Sep 24, 2007 12:24 pm Post subject:

Mike G wrote:
Also thanks for reminding of Nick S's work a few weeks back; I was heavily curious where he would take that.
Thanks for reminding me that I'm not forgotten. I have been meaning to do some additional work on that before the season starts, but work keeps interfering. My argument for a system based on "player wins" adding up to expected wins rather than actual wins was made here and was relatively simple: "The argument for using "expected" is that it keeps everything on the same "scale". So two teams with the same defensive eFG% will get credit for the same number of "wins" from their eFG% defense." "The other reason is that if someone calculates this in-season, expected wins are probably a better predictor of future success than actual wins particularly over a small sample size (in other words, if you calculate the numbers after 10 games, and two teams are both 6-4 but one team has a +8 point differential and the other team has a +0 point differential, the differential is a more accurate measure of ability than record)" I would add to that the idea that expected wins is a better way to compare the strength of teams historically and, therefore, a better way to compare the performance of players from different seasons.
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Mon Sep 24, 2007 12:36 pm Post subject:

Ok Mike, I understand your perspective. I've shared my perspective about adding individual shot defense enough. DCS has it. I will look to that more and maybe create my own version of it later.
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Mon Sep 24, 2007 1:06 pm Post subject:

Well Mike, EWins goes primetime with Henry Abbott's mention of it and link to this board about it this morning. If folks haven't seen it yet: http://myespn.go.com/nba/truehoop Dan and David's presentation gets more publicity and discussion too. (And in a bankshot WARP as presented by DW might get some exposure.) P.S. I found this formula in the mentioned player ranking thread that covers EWins T = Sco + .98*Reb + 1.335*Ast - .25*PF + 1.5*Stl - 1.5*TO + 1.75*Blk Still accurate?
Back to top

mateo82



Joined: 06 Aug 2005
Posts: 211
Posted: Mon Sep 24, 2007 5:02 pm Post subject:

Well, I would think it would have to measure expected wins, since the difference between expected wins and actual wins is essential chance
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3615
Location: Hendersonville, NC
Posted: Tue Sep 25, 2007 7:10 am Post subject:

mateo82 wrote:
Well, I would think it would have to measure expected wins, since the difference between expected wins and actual wins is essential chance
Chance, or 'tanking'. How 'bout them Celtics? -- shoulda won 32 games but only won 24. WS total 24 (X3), while PW total 32. The difference is significant. Mtn, those are one version of the weights to arrive at T. In playoffs, they can change quite a bit; for seasons, pretty close to what you have (1.5 for blocks is more typical, and .99 for Reb.) The weights change a bit for 'best fit' between wins and eWins. Other parameters also change, within some limits._________________` 36% of all statistics are wrong
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Tue Sep 25, 2007 12:31 pm Post subject:

Thanks for sharing Mike. It will take some time to look at. If I have any observations or questions that might be worthwhile I'll pass them along later.



Page 2

Author Message Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Tue Sep 25, 2007 3:22 pm Post subject:

On EWins formula my initial reaction is assists weight is higher than my past practice / prejudice but I'll have to go back and look at all the measures of assist impact and reassess. The spreadsheet is large with many formulas and I certainly haven't mastered them (and may not) ... but I did a quick compilation of EWins for 6 main Sonics last season and put them in a table with WinShare scores at basketball-reference.com and Wins Produced. To normalize them I pro-rated the EWins and WinsProduced scores from their expected wins basis to a pro-rated version of actual wins (ok or disagree?). WinShares are based on actual wins but I divided by 3 to get Wins. Wins Produced allows negative scores and that helps account for its higher scroes for the best players, at least partially. For what it might be worth to the discussion I present this data: prorated to actual W WShares WProd EWins Average allen,ray 5.3 8.1 6.2 6.6 lewis,rashard 6.3 7.8 6.3 6.8 wilcox,chris 5.3 5.3 5.5 5.4 ridnour,luke 2.0 1.9 2.9 2.3 collison,nick 4.0 5.1 3.7 4.3 watson,earl 1.7 3.2 2.5 2.4 Score/Avg for blend WShares WProd EWins allen,ray 0.81 1.24 0.95 lewis,rashard 0.93 1.15 0.93 wilcox,chris 0.99 0.99 1.02 ridnour,luke 0.88 0.83 1.29 collison,nick 0.93 1.20 0.87 watson,earl 0.68 1.30 1.02 Anyone have comments or questions about these results? Obviously a full league correlation study amongst these (and possibly others) could be done to try to identify why they move the way they do and in relation to each other. Going multi-year would add robustness. But at the surface they track fairly closely and in line with my subjective guesses before I looked up the values. The existing tools may be pretty good for first cut purposes.Last edited by Mountain on Wed Sep 26, 2007 12:15 pm; edited 1 time in total
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3481
Location: Hendersonville, NC
Posted: Wed Sep 26, 2007 6:13 am Post subject:

I do weigh Ast rate more heavily than Sco or Reb rates. This is made possible (still makes better correlation between eW and PythW) by a couple of factors: -- There's an estimate for players % of FG unassisted. Players' Sco rates are boosted (or depressed) by whether this % is more or less than average. -- Ast rates are moved by the factor (TmPts/OppPts)^1.5 This 2nd factor makes Ridnour's Ast rate 5.61, rather than 5.69; and he has .04 fewer eWins on the year. Assists are important as a proxy for lots of essential actions which do not become boxscore stats: bringing the ball up, spacing the floor, and of course all the passes that aren't assists (because the receiver either passed or missed)._________________` 36% of all statistics are wrong
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Thu Dec 11, 2008 12:38 pm Post subject:

How do the top teams look on eWins distribution by player or by minutes weighted position? Ever calculate regular season league averages for teams or just playoff teams or top 4 or recent champion averages? Or for scoring, rebounding and assist rates? How about similar distributions for other metrics? Any advocate / user of them want to share these?
Back to top

jkubatko



Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH
Posted: Thu Dec 11, 2008 2:33 pm Post subject:

Wow, resurrecting a thread that's been dead for almost 15 months. That might be a record or something._________________Regards, Justin Kubatko Basketball-Reference.com
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Thu Dec 11, 2008 3:46 pm Post subject:

Nope. Now this thread didn't need to be resurrected but I went looking for old threads because there is good stuff in the archives worth recalling and eWins is a worthy topic folks occasionally ask for more background on and in this case the title fit with my comments so I just went with it. In what other field of research would referencing a conversation from 15 months ago be treated as unusual or unwelcome? Is always staying in the moment and the chatter of the few days being argued as preferable / better?
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3481
Location: Hendersonville, NC
Posted: Fri Dec 12, 2008 7:40 am Post subject:

Mountain wrote:
How do the top teams look on eWins distribution by player or by minutes weighted position? Ever calculate regular season league averages for teams or just playoff teams or top 4 or recent champion averages? Or for scoring, rebounding and assist rates?
eWins gives credit to players, for their team's wins. A team's eWins can be guessed by the formula: eW = (pW + 41)/2 - where pW is pythagorean. Ranking teams by Sco, Reb, and Ast could be done by checking their PtDiff, Reb% and Ast%, I think. I've never used team rates in those categories. You can download this year's file and play with it, sort by position, etc. Not sure what you're asking about, at the top._________________` 36% of all statistics are wrong
Back to top

Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Fri Dec 12, 2008 11:06 am Post subject:

Refresh my memory, what's the rationale behind this part? Quote:
eW = (pW + 41)/2
Shouldn't eW = pW? Or eW = W?
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Fri Dec 12, 2008 3:11 pm Post subject:

I am basically wondering about team structure or DNA as I have in the past with types but now thinking about it in terms of impact independent of style of play. Do the top teams have a certain. frequently similar distribution of eWins contributed by players down the ranks from the stars to the first line bench guys? I think it is also worth checking if there is a similar distribution for successful teams by where the eWin contributors actually play, where the major impacts come from by position. Calculating regular season league averages for teams or just playoff teams or top 4 or recent champion averages would give information to assess whether a certain type of eWin contribution or position pattern where the contributions come from appears to matter. Looking at scoring, rebounding and assist rates would be taking that pattern analysis / recognition exercise deeper into the details. Maybe the structure of impact thru the rotation or position don't matter or matter much. Maybe it is just get edge wherever you can. But I'd guess that probably the structure of today's game will give an edge to some designs over others, so why not look and try to understand the patterns better than anybody else and make your best educated guess on the pattern to choose or tweak what you've got and can get it in that direction as far as you can. This could and should (in my mind) by done with offensive/defensive adjusted splits and net counterpart data and could be done for other box-score based metrics too if you prefer one of them to eWins or just want to study and see as much as you can. Maybe I'll work thru it using your file Mike, maybe not. Checking what you've done and throwing the idea out there for any interested. I assume this type exercise gets done in the best analytic shops. I would do it if I had that role. The distribution of contributions has to be considered in the context of 5 man lineups and the rotation. That is where the rubber meets the road. So then you get back to needing the adjusted 5 man lineup data I advocated some for, Eli produced a first cut of before getting snapped up and brought inside. Roland at 82 games & with a team and perhaps other insiders may have it as well. That is probably the most important data out there or out there to get for this team game. Or at least the most important supplement to the rest of what is out there in public. How important is context for impact, all told? I assume it is more than enough to make the difference between winning a title and losing the conference finals. If so then it is vital. I don't think just summing individual stats produces essentially the team result or is the sufficient projection / planning approach. The last few (or even more than a few) points influenced by context demand context based analytics.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3481
Location: Hendersonville, NC
Posted: Sun Dec 14, 2008 5:05 am Post subject:

davis21wylie2121 wrote:
Refresh my memory, what's the rationale behind this part? Quote:
eW = (pW + 41)/2
Shouldn't eW = pW? Or eW = W?
No, eW = Equivalent Wins a player adds to a hypothetical average team. The rationale is that when I add a team's player eWins, it just happens that pW = 2*eW - 41 One system that ties player wins to team wins is Win Shares. Recall that players for the Celtics last year were all on bad teams the year before. Here are some WS and eW, per 484 minutes; and a 'prediction' based on their actual 2008 minutes x their 2007 win-producing rates: Code:
eWins/484 eW? Min WS/484 WS? Celtics 2007 2008 Pre08 (2008) 2007 2008 Pre08 Pierce 1.86 1.85 11.1 2874 1.50 2.05 8.9 Allen,R 1.62 1.31 8.8 2624 1.37 1.75 7.4 Garnett 2.22 2.50 10.7 2328 1.70 2.62 8.2 Rondo 0.74 1.24 3.5 2306 0.63 1.49 3.0 Perkins 0.45 1.14 1.8 1912 0.43 1.54 1.7 totals: 35.8 12044 29.3
eWins says the youngsters improved dramatically, while the vets were pretty much the same (Ray cancelling KG). WS says everyone improved stupendously, in their win-producing attributes. Of course, in retrospect we know the defense improved a lot, chemistry was superb, etc. But without knowing those specifics, we can look at the totals at the bottom. Note that the minutes has about 7000 unaccounted for in this short list. eWins (predicted) is 6.5 more than WS. This is really as much as 13 more wins expected (as in 2*eW-41) from just these 5 players, without any change in previous win rates. Because WS are depressed for players on bad teams, there's no built-in way to predict wins when these players are gathered together to form a much better team -- is there?_________________` 36% of all statistics are wrong
Back to top

Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Sun Dec 14, 2008 1:41 pm Post subject:

I think you're confusing ability with value again... The "wins a player adds to a hypothetical average team" would fall under the category of ability, an attempt to completely remove the player from his actual context and drop him into a different context. So in this sense, eWins is apparently trying to measure ability. Win Shares wants to measure value -- and the sum of a team's players' "wins added" must be tied to the actual # of wins (real or pythagorean) the team produced in reality. In a value measure, it makes absolutely no sense to artificially allocate more wins to a bad team's players (or less to a good team's ones) than the team actually won in real life. The whole must (roughly) equal the sum of its parts. So, as it turns out, comparing eW to WS is like comparing apples and oranges. And I don't think I realized this until right now.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3481
Location: Hendersonville, NC
Posted: Sun Dec 14, 2008 2:11 pm Post subject:

davis21wylie2121 wrote:
.. an attempt to completely remove the player from his actual context and drop him into a different context..
Actually, the opposite. E-Wins isn't tied to context, but is a universal expression of Value, independent of context. If a player's value is only relative to where he played in a given season (or part-season), what does that tell us about his value? Not much, I think. If you drive a car across the Plains and get 35 mpg -- and it's known that you had a tailwind averaging 20 mph -- wouldn't it be worth knowing what your car's mpg is in the absence of wind? Or into a 20 mpg headwind? The fleeting 'value' of your high-mpg car might be a useful quantity to quote, if you're trying to sell your car. But by accounting for the context (of the tailwind), one gets a more accurate description. A role player on a great team may shop himself around the league and get paid more by a lesser team. Does his 'value' decrease, if equal production creates fewer WS? Team eWins rank teams in just the same order as PythWins rank them. There's no artificial adjustment made. Player wins don't sum to team wins. The thread starts with these inquiries. The whole does equal the sum, in context._________________` 36% of all statistics are wrong
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3481
Location: Hendersonville, NC
Posted: Sun Dec 14, 2008 2:26 pm Post subject:

Try this thought experiment. You and I and 2 other guys are playing games of 2-on-2. We are all equal players (ability, value), and we split 20 games, 10-10. We all have essentially equal WS, eW, whatever: Just about 5.0 apiece. Then your teammate gets hurt. To continue our 'league', you draft a guy who is just a stiff. He can do very little to contribute. We double team you, and make him miss his occasional shot by yelling at him. Now it's very difficult for you to win a game. You may win one of 20, because we were hung-over perhaps. In 20 games, you may get 20% more points, taking 80% more shots; double your turnovers, etc. By most measures you've been more productive and less efficient. But eWins says you're the same player you have been. Yet, your win shares have plunged to nearly zero. Some ratings would have you negative. Has your 'value' gone to near-zero? Your team might total 5.5 eW in 20 games, to my team's 14.5 . eWins : 5.5*2 - 10 = 1.0 In other words, by just maintaining your productivity, with a teammate about 1/10 as productive, your team is expected to win 1 game in 20. And your 'value' is constant; as (I think) it should be._________________` 36% of all statistics are wrong
Back to top

jkubatko



Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH
Posted: Sun Dec 14, 2008 2:27 pm Post subject:

Mike G wrote:
One system that ties player wins to team wins is Win Shares. Recall that players for the Celtics last year were all on bad teams the year before. Here are some WS and eW, per 484 minutes; and a 'prediction' based on their actual 2008 minutes x their 2007 win-producing rates: Code:
eWins/484 eW? Min WS/484 WS? Celtics 2007 2008 Pre08 (2008) 2007 2008 Pre08 Pierce 1.86 1.85 11.1 2874 1.50 2.05 8.9 Allen,R 1.62 1.31 8.8 2624 1.37 1.75 7.4 Garnett 2.22 2.50 10.7 2328 1.70 2.62 8.2 Rondo 0.74 1.24 3.5 2306 0.63 1.49 3.0 Perkins 0.45 1.14 1.8 1912 0.43 1.54 1.7 totals: 35.8 12044 29.3

Wait a second... You've already said pW = 2 * eW - 41. So using the numbers above, doesn't that mean that eW suggests that this team would be projected to win at least (2 * 35.8) - 41 = 30.6 games? And isn't that very close to the Win Shares projection of 29.3 wins?_________________Regards, Justin Kubatko Basketball-Reference.com
Back to top

Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Sun Dec 14, 2008 2:34 pm Post subject:

This is what I'm trying to say: There is no such thing as a "universal expression of Value". That's called "Ability". Value embraces context (if a player is on a bad team, so be it), while ability rejects it. I hate to break it to you, but if you're attempting to show how many wins a player produced on a theoretical average team, that's an ability metric, not a value metric, because you've tried to remove a player from his actual context.
Back to top

Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Sun Dec 14, 2008 2:49 pm Post subject:

If I may quote Ed Kupfer for a moment: Ed Küpfer wrote:
Imagine a two-on-two tournament. The favoured team is comprised of one mediocre player and one player so great, he ranks among the best of all time. Now imagine that every opponent is determined not to let the great player beat them, that they will double-team him continuously and force the mediocre player to shoot. Now, because of the double-team, the mediocre player has no one defending him, and scores pretty much at will. Opponents, still fearing the great player's awesome talent, stay with the double-teams, and keep hoping that the mediocre will miss, which is unlikely since they aren't really defending him. Now, regardless of how the team fared in the tournament, the stats will show that the mediocre player contributed far more value than the great player, becuase, as it turned out, he was the one taking—and making—all the shots. Our stats will give the mediocre player the MVP. This is the way it should be. Just remember that most valuable does not mean most talented.
Pertaining to this discussion, eW is apparently trying to measure talent; WS is trying to measure literal value.

Re: eWins and other Individual Wins accrediting systems

Posted: Fri Apr 15, 2011 9:42 am
by Crow
Author Message Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Sun Dec 14, 2008 3:15 pm Post subject:

Btw, the full "eWins formula" isn't published, is it? I tried to reverse-engineer it from your spreadsheet once, but every formula always came back to an unexplained constant. You've hinted at things here and there, but I've never seen the full step-by-step explanation... Win Shares, on the other hand, is completely transparent -- anyone can read the steps and know exactly how it's calculated. So I'd like to ask you, Mike -- if you're willing -- to do the same. Lay it out step-by-step: Exactly how is T-rate calculated? How were the weights obtained? How do you go from T-rate to eWins? Until I see exactly how it works, I think this is another reason why it's unfair to compare eW to WS.
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Sun Dec 14, 2008 3:21 pm Post subject:

Maybe I haven't thought about the value and ability terms long enough and precisely enough or fully processed the rest of the discussion but I will try this side add to the discussion: What does last season's Boston team formation and results tell us? 1. Maybe that Pierce's defensive ability was undevalued by a team level defense metric because of the bad teammate context? His counterpart defense data was good and I believe (in contrast to others) it is often correctly indicative of defensive performance at the level of good, average, poor (not always) and is worth noting and can be a building block. 2. That Allens's counterpart offense transferred as Mike's and other metric can estimate to one degree of accuracy or another. And that his counterpart defense wasn't as bad as his team based defensive marks and that he could fit well enough in a really good defense context. Hard to say directly from "Defensive Rating" how much of the team defensive failure in Seattle was really Allen. Now adjusted defensive +/- says he was a negative but by playing him heavily with Garnett they got by and he actually had a better raw opponents scored while on the court than Garnett. P.S. By counterpart defensive data Allen is playing like a defensive player of the year candidate this season- 41% eFG% allowed for a PER9 at SG and even better at SF. Maybe he is a good student - and a proud champion- and now that he knows how to play good defense and has reaped its rewards he is doing it. Though may this is exaggerated in some fashion. Early help to avoid problems? Tony Allen is doing it too so maybe Boston's system is trying especially hard to take away the SG opponent? 3. That Garnett's team defensive mark based heavily on his help defense transferred along with his individual defense. 4. That Boston's defensive system under Thibodeau's leadership is a key factor. Other managers could have seen lesser return. I guess that can be true for offense or defense, net counterpart level impacts or team. So maybe this is just a team specific success story rather than illuminating in general about impact transfer mechanics or efficiencies. Or the right metrics to capture that information. I don't know. Maybe I'll think about it more later.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3527
Location: Hendersonville, NC
Posted: Sun Dec 14, 2008 5:41 pm Post subject:

OK. Last time I checked b-r.com's Glossary, in some places it says, "Read BoP", doesn't it? In reverse, eW = (T-Tr)*Min/X - T is the aforementioned player T rate (weighted sum of boxscore stats, standardized) - Tr is 'replacement level' T - Min is minutes played - X is a constant selected by Excel's GoalSeek function. It causes league eWins to total league games played. (In playoffs, perhaps just a playoff series.) Currently, X = 5381 and Tr = 12.09 Tr is one of several variables that are tweaked up and down to achieve the closest fit between avg team pW and xW. - pW are pythagorean, from points for/against - xW (expected W) = eW*2 - G/2 The categorical stats and weights -- right now, early in the season, they may vary from their normal values -- are: Sco * 1.0 (by definition, though this category has major standardizations applied) Reb * 1.0 Ast * 1.33 PF * .25 Stl * 1.35 TO * -1.3 Blk * 1.15 A 'minutes factor' is applied to all categories. It is: fac = (mpg/36) * (36/mpg)^a * Sta - 'a' is a variable, currently .195 - Sta is a 'vs-starter factor': St%^b - - St% is estimated fraction of starters faced - - exponent 'b' another variable, currently .165 PF, Stl, TO, and Blk are merely their per-game average * fac. Ast = APG * fac * TmFG/G/32 * (TmPPG/OpPPG)^c * AsBi^d - exp 'c' is at its 'ceiling' of 1.20; sort of a 'cheat factor' - AsBi is Home Assist Bias, the average of last year and this year; derivation can be found in another thread. - - exp. 'd' currently = .25 Reb = RPG * fac * r/(TmRPG + OpRPG) * (2*TmRb%)^f - 'r' is arbitrary avg Tm+Op Reb/G = 88 - 'f' is currently 1.000 - TmRb% is (my own version of) = (OReb% + DReb%)/2 Sco = PPG * fac * UAFac * (100/OpPPG)^j * (TS%/k)^m - UAFac is 'unassisted points' (estimated: see yet another thread) - exp j is another 'cheat', also at its ceiling (1.10) - k is .525 - exp. 'm' is generally around .5, now .55 I've left out a bit, and it's still exhausting. I'm not married to many of these factors. It's really just a demonstration project, and I'd be delighted if someone can streamline it, keep some basics, and replicate the nice fit between xW and pW. I get an average 'error' of .47 (after 21-24 games); the avg 'error' between W and pW is 1.13 ._________________` 36% of all statistics are wrong
Back to top

DLew



Joined: 13 Nov 2006
Posts: 222
Posted: Sun Dec 14, 2008 6:34 pm Post subject:

I'd like to defend Mike's preference for a system that rates ability. To me in many cases a system that removes context and puts all players on a level, comparable, playing field is more useful than one that simply evaluates player value within context. I'm not saying that I'm sold that eWins does this, or that it is better than WinShares, just that I understand, and am sympathetic to, Mike's ability-based view point.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3527
Location: Hendersonville, NC
Posted: Sun Dec 14, 2008 6:42 pm Post subject:

jkubatko wrote:
Mike G wrote:
Code:
eWins/484 eW? Min WS/484 WS? Celtics 2007 2008 Pre08 (2008) 2007 2008 Pre08 Pierce 1.86 1.85 11.1 2874 1.50 2.05 8.9 Allen,R 1.62 1.31 8.8 2624 1.37 1.75 7.4 Garnett 2.22 2.50 10.7 2328 1.70 2.62 8.2 Rondo 0.74 1.24 3.5 2306 0.63 1.49 3.0 Perkins 0.45 1.14 1.8 1912 0.43 1.54 1.7 totals: 35.8 12044 29.3

Wait a second... You've already said pW = 2 * eW - 41. So using the numbers above, doesn't that mean that eW suggests that this team would be projected to win at least (2 * 35.8) - 41 = 30.6 games? And isn't that very close to the Win Shares projection of 29.3 wins?
So it is. But 12044 minutes is about 7700 short of a full team-season. If those 7700 minutes were taken by replacement players, then yes, the Celts would win about 29 G. As it turned out, the starters totalled 40.5, and everyone else 13.4, for a total of 53.9. 53.9 * 2 - 41 = 66.8 (They won 66, pyth = 66.3)_________________` 36% of all statistics are wrong
Back to top

Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Sun Dec 14, 2008 7:42 pm Post subject:

Thank you, Mike. It's going to take a while to unravel that post, and I'm still not clear on where some of your variables and exponents are coming from (and why anything is raised to an exponent in the first place), but at least it's a start toward understanding what makes eWins tick. As for the WS description, it does in fact say "Read BoP" in various places. At which time you go to the bookstore and pick up a copy of BoP, which outlines Dean's methods in excruciating detail (I thought you already had a copy anyway?). The point is, every step of the Win Shares calculation is out there publicly if someone wants to replicate it themselves. Before today (and perhaps even still), I couldn't say the same of eWins. And DLew, I agree that there's definitely a place out there for an ability/talent rating. I just think eWins is a bit misleading in the sense that it looks like a literal-value metric and is spoken about in that way, when in fact it is an ability metric, or at the very least a performance metric. Value doesn't adjust for context, because in many cases context determines value.
Back to top

jkubatko



Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH
Posted: Mon Dec 15, 2008 12:24 am Post subject:

Mike G wrote:
So it is. But 12044 minutes is about 7700 short of a full team-season. If those 7700 minutes were taken by replacement players, then yes, the Celts would win about 29 G. As it turned out, the starters totalled 40.5, and everyone else 13.4, for a total of 53.9. 53.9 * 2 - 41 = 66.8 (They won 66, pyth = 66.3)
So, in the example that you cherry-picked to show how superior eWins is to Win Shares, you have conceded that both systems show that the top five players would have projected to account for 29-30 wins based on the previous season's rates. Correct? If so, how does this demonstrate the pure awesomeness of eWins?_________________Regards, Justin Kubatko Basketball-Reference.com
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Mon Dec 15, 2008 3:31 am Post subject:

So am I understanding correctly that an offensive and defensive rebound would contribute the same amount to rebounding rate? With some of your weights giving rather large credit for certain actions, presumably because of what they lead to or indicate about nonboxscore performance, I wonder how closely your ratings align with adjusted +/= or how a blend of adjusted +/- and your rating serving the role of a statistical +/- would do as a projection tool.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3527
Location: Hendersonville, NC
Posted: Mon Dec 15, 2008 6:25 am Post subject:

jkubatko wrote:
... both systems show that the top five players would have projected to account for 29-30 wins based on the previous season's rates. Correct? If so, how does this demonstrate the pure awesomeness of eWins?
Well, let's deal with a full deck. Throw in Posey, House, Powe, TAllen, and Scalabrine. For the top 10 Celtics, eW*Min totals 45.1 and WS 42.5 Davis, Cassell, PJ, Pollard, and Pruitt came along to add -- after the fact -- another 2.4 eW. Supposing we 'knew' this beforehand, we could add a similar amount to WS: eW 47.5 - WS 45.0 Now, here's the awesome part of eWins: xW = 47.5*2 - 41 = 54 The difference between 54-28 and 45-37 is pretty substantial. The similarity (29-30 wins) when zero-win teammates are alongside may be worth knowing too, but that's not a very realistic team scenario. Even this year's Thunders have more than 5 >replacement players. Like, six. I will note that WS more distinctly predicted the punch that Boston's bench would have last year. Low-usage, high-efficiency guys (Posey, notably) have greater 'value' when they are role players alongside stars?_________________` 36% of all statistics are wrong
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3527
Location: Hendersonville, NC
Posted: Mon Dec 15, 2008 6:43 am Post subject:

DLew wrote:
I'd like to defend Mike's preference for a system that rates ability. To me in many cases a system that removes context and puts all players on a level, comparable, playing field is more useful than one that simply evaluates player value within context. I'm not saying that I'm sold that eWins does this, or that it is better than WinShares, just that I understand, and am sympathetic to, Mike's ability-based view point.
Maybe one system has more value, and the other has more ability? Appreciate the appreciation, though. eWins strikes out into the gulf between win-differential and point-differential._________________` 36% of all statistics are wrong
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3527
Location: Hendersonville, NC
Posted: Mon Dec 15, 2008 7:11 am Post subject:

Mountain wrote:
.. Hard to say directly from "Defensive Rating" how much of the team defensive failure in Seattle was really Allen.
Ray's career DWS rate was .44 of avg before coming to Boston; since then, it's 1.44 . Quote:
... an offensive and defensive rebound would contribute the same amount to rebounding rate?
Yes and no. Player rebounds are scaled to team Reb%, which I calculate by averaging OReb% and DReb%. (These estimate missed FT and remove them from the calc.) Effectively, this averaging boosts the value of the OReb (at a team level), since it counts those 25-30% as heavily as the other 70-75% (DReb). As you suspect, most stats include a bonus or penalty for 'uncounted' actions. On a given play, an OReb has great value; but missing puts you in bad defensive transition. I have tried to find a value difference between O and D Reb, but there doesn't seem to be any. It is a kind of 'retro +/-', since all these things which go into Pt-Diff are adjusted to 'best fit'. I don't call it regression; it's more art than science._________________` 36% of all statistics are wrong
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3527
Location: Hendersonville, NC
Posted: Mon Dec 15, 2008 9:25 am Post subject:

davis21wylie2121 wrote:
.. It's going to take a while to unravel that post, and I'm still not clear on where some of your variables and exponents are coming from (and why anything is raised to an exponent in the first place), but at least it's a start toward understanding what makes eWins tick. ..
y = mx+b creates a straight-line relationship, with 2 variables. y = x^c creates a curvy one, with 1 variable. Fewer is better? I should stress that there are a couple of discussions going on here. The complicated one is regarding the T rate, which could be debated forever. I think T produces a pretty non-controversial result for the most part. It says LeBron is the best, then Wade and Paul, then Dwight and Duncan; etc. The 2nd, much more interesting, and infinitely simpler, is the relation of eWins to xWins. It's all about losing the idea that value = team wins. Does a 60-win team have value equal to 2 or 3 teams that total 60 wins? Does a player's real value increase when he goes to a better team and does less? What else in the real world supports this definition of value? To create a more complete protein, eat beans and rice together. Eating beans for lunch and rice for dinner will create less protein inside you. Together, value is optimized. If you have no beans, does the rice have less value? You will get less protein; but if the alternative is to starve, your rice could be said to have more value, not less. A team that's starved for wins, or for an entertaining product, gets value from it's players. Is value really just about the number of wins currently accrued? If their superstar comes back from injury, do his teammates have more value as wins roll in? Intuitively, it seems the opposite is true: the other guys become more expendable. Shouldn't there be even a possibility of 'absolute value' that is equivalent to performance (ability)?_________________` 36% of all statistics are wrong
Back to top

gabefarkas



Joined: 31 Dec 2004
Posts: 1312
Location: Durham, NC
Posted: Mon Dec 15, 2008 10:17 am Post subject:

Mike G wrote:
A role player on a great team may shop himself around the league and get paid more by a lesser team. Does his 'value' decrease, if equal production creates fewer WS?
Possibly. Or it might mean you previously had overvalued him while on the great team.
Back to top

gabefarkas



Joined: 31 Dec 2004
Posts: 1312
Location: Durham, NC
Posted: Mon Dec 15, 2008 10:24 am Post subject:

davis21wylie2121 wrote:
Thank you, Mike. It's going to take a while to unravel that post, and I'm still not clear on where some of your variables and exponents are coming from (and why anything is raised to an exponent in the first place), but at least it's a start toward understanding what makes eWins tick.
What's wrong with using exponents? If that's how the data looks, that's how you should deal with it. Do you think everything is normally distributed? Have you ever looked at a histogram of PERs of all players in a given season?
Back to top

gabefarkas



Joined: 31 Dec 2004
Posts: 1312
Location: Durham, NC
Posted: Mon Dec 15, 2008 10:37 am Post subject:

Mike G wrote:
davis21wylie2121 wrote:
.. It's going to take a while to unravel that post, and I'm still not clear on where some of your variables and exponents are coming from (and why anything is raised to an exponent in the first place), but at least it's a start toward understanding what makes eWins tick. ..
y = mx+b creates a straight-line relationship, with 2 variables. y = x^c creates a curvy one, with 1 variable. Fewer is better?
Those each have two variables: x and y. However, the linear equation has two constants (or coefficients, if you'd like): m and b. On the other hand, the power equation, as you've written it, only has one constant: c. However, it's not a fair comparison, since you're using a very specific version of the exponential equation in which you've assumed a lot of other constants are equal to 1. It would be like saying the linear equation is just "y = mx" with the assumption that b=0. The generalized exponential distribution is much more complex, with a few more coefficients involved.


Author Message Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Mon Dec 15, 2008 11:32 am Post subject:

gabefarkas wrote:
What's wrong with using exponents? If that's how the data looks, that's how you should deal with it.
There's nothing wrong with using exponents, when appropriate. I just want to make sure this is one of those cases. --Last edited by Neil Paine on Mon Dec 15, 2008 11:56 am; edited 1 time in total
Back to top

Neil Paine



Joined: 13 Oct 2005
Posts: 774
Location: Atlanta, GA
Posted: Mon Dec 15, 2008 11:55 am Post subject:

Mike G wrote:
I should stress that there are a couple of discussions going on here.
True. The philosophical argument is a tired one which can never be definitively answered, because it's essentially going to boil down to an opinion. So instead of fighting that battle again, I just want to know where the weights in T-rate come from -- did you run a regression? How do you determine the exponents? If I sat down with this year's raw stats in excel, how would I build T-rate? You started answering that question above, but I still don't know how some of the variables are determined.Last edited by Neil Paine on Mon Dec 15, 2008 11:55 am; edited 1 time in total
Back to top

Mountain



Joined: 13 Mar 2007
Posts: 1527
Posted: Mon Dec 15, 2008 11:55 am Post subject:

"Ray's career DWS rate was .44 of avg before coming to Boston; since then, it's 1.44." That this big a change can happen suggests to me that he isn't as poor as the former rating or as good as the latter rating and that Defensive Win Shares isn't capturing defensive ability. It is assigning defensive value based on team level results (that is the intention) but that is going to be based on the 5 man product, 80% of which or so isn't that player himself. Hence I give some attention to the counterpart defensive data to check 1 on 1 ability that might transfer between teams.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3605
Location: Hendersonville, NC
Posted: Mon Dec 15, 2008 12:58 pm Post subject:

An exponent seems to be appropriate when a factor is close to 1 : - (TS%/.525)^a : factor applied to scoring rate - (TmReb/OpReb)^b : factor applied to rebound rate - (TmAssistBias)^c : applied to assist rate - (%StarterFactor)^d : applied to all rates When the exponent approaches zero, the factor approaches 1: it's a non-factor. When the exponent goes to 1, the factor is the direct ratio or value. What do you call a variable constant -- a parameter?_________________` 36% of all statistics are wrong
Back to top

Harold Almonte



Joined: 04 Aug 2006
Posts: 616
Posted: Tue Dec 16, 2008 5:23 pm Post subject:

Mike. Aren't there negative eWins, or eLosses, or negative players? What 's the maximum unproductivity in this metric? I understand the difference between to win or loose a game is a certain, and sometimes minimmal point differential (sometimes even undeserved). But this is the point differential between teams, every team (and every player) has their own "performance point differential", and one of them is negative (long or short but negative). Win is a dialectic cathegory attached to Lose, that's what I see it. Your metric is a no pythagorean win metric, it is more like a "race to wins", with first prize, second prize, etc.. , the loses are obligued to be put just because one need to substract from 82. The own players-teams's race to break the even between having the "point differential" in favour, or against, is another thing. Wow, too many words to say this is a metric with no points allowed. It can measure some value and levels of value, but not almost all the unvalue.
Back to top

Mike G



Joined: 14 Jan 2005
Posts: 3605
Location: Hendersonville, NC
Posted: Tue Dec 16, 2008 6:20 pm Post subject:

Harold Almonte wrote:
... It can measure some value and levels of value, but not almost all the unvalue.
I could not have said it better. After 359 games, NBA players net 3.0 negative eWins. Ricky Davis 'leads' at -.21 That's just .008 of the total, but pooled from .23 of NBA players. Last year, about half the league's players (230, or 7.6 per team) accounted for 90% of the wins. By the end of the year, players in the negative totalled -2.3 eW ('led' by Jason Collins' -.5), not quite .002 of the total (1230). Season negatives were only .07 of all who played. Apparently, as the year goes on, sub-replacement play tends to subside. If a guy is really that bad, he goes back to the bench (or the D-League); if he isn't that bad, he gets better. A quarter-season has 4x the fraction of 'losing' play, that a full season has. Presumably, a single game has a much greater fraction. Single possessions, greater still. Eventually, it's an order of magnitude._________________` 36% of all statistics are wrong
Back to top