Point Allocation
Copyright Iain Fyffe, 2002
For those who may not know, Bill James is quite a brilliant man. He's known primarily, of course, for his work as a stathead in baseball. It has become fashionable of late (especially amongst younger statheads) to decry James' work. I'll not get into that; I'll just say this: his pure writing about baseball is arguably more impressive and engaging than his statistical work about baseball. Even if he wasn't a brilliant statistician, his work would still be invaluable to any fan who considers himself to be knowledgable about the game.
Fortunately for us hockey folk, some of James' work and ideas can be translated for use in hockey, or can at least be used for inspiration. This paper describes the development of the Point Allocation system, which is a method of evaluating players based on their contributions to their team's success (or lack thereof). It is based on two bits of Bill James; I have adapted his Marginal Runs analysis, which forms the basis for his Win Shares system of player evaluation, and I have extrapolated quite a bit from a fairly casual remark he made in The Politics of Glory.
I'll start with that relatively innocuous comment. James was discussing the common assertion that defence is not reflected well in statistics (at least, the statistics that most people talk about). He pointed out that, to a degree, defence is in fact reflected, through the length of the player's career, and the amount that he plays. For instance, Brooks Robinson was not a particularly great hitter. And yet, he had an exceptionally long career. Why? You know the answer: he was probably the greatest defensive third baseman who ever lived.
Epiphany! Look at this: Guy Carbonneau and Bob Gainey (for example), both with unimpressive offensive totals, both with long careers. Both renowned as defensive players, even if the stats (like plus-minus) "don't show it".
Epiphany again! Maybe we can take this concept down to a team-season level. That is, say we have two players on the same team, who contribute the same offensively, on a per-minute basis. One player plays 15 minutes per game, the other plays 18 minutes per game. Since these players are offensive equals, there can be only one explanation for the discrepancy in playing time: defence. That second player's defence must be sufficiently better than the first's to warrant three extra minutes of playing time per game. More on this later.
Marginal Goals
Now we move on to the basic ideas behind James' Win Shares system, which I have adapted to create the Point Allocation system. Win Shares is a way of distributing a team's wins amongst its players, based on their relative contributions to the team's success. The building block of Win Shares is Marginal Runs; therefore the building block for Point Allocation is Marginal Goals.
James discovered a new method that predicts team success similarly to his famous Pythagorean analysis (which itself has been adapted to hockey by Marc Foster). I'll explain it in hockey terms. The following formula is an excellent predictor of a team's winning percentage:
E(Pct) = (MGF + MGS) / (2 x AvgG)
Where E(Pct) is expected winning percentage; MGF is Marginal Goals For, calculated as the team's goals for less one-half the league-average goals per team; MGS is Marginal Goals Saved, calculated as one and one-half times the league- average goals per team less the team's goals against; and AvgG is the league-average goals per team.
Marginal Goals is no better at predicting winning percentage than Pythagorean analysis; in fact, it's probably slightly worse. However, what Marginal Goals allows us to do is apportion the team's winning percentage (in the form of points) between a team's offence and defence, as follows:
OP = MGF / TMG x Pts
DP = MGS / TMG x Pts
Where OP is offensive points (points attributable to offence); DP is defensive points (points attributable to defence); TMG is Total Marginal Goals (Marginal Goals For plus Marginal Goals Saved); Pts is the team's points (ties plus two times wins); and MGF and MGS are defined as above.
We simply cannot do this with Pythagorean analysis. Say we have two .500 teams in a league where 300 goals is average. Team A scores and allows 350 goals, while Team B scores and allows 250. Pythagorean analysis will tell us that both of these teams should both be at .500, which they are. But Team A's success is clearly tied to its offence, while Team B relies more on defence. Marginal Goal analysis allows us to determine how much success is attributable to offence and defence (in this example, Team A is 67% offence and 33% defence, and Team B is 33% offence and 67% defence).
Now if you're wondering why 0.5 and 1.5 are used, rather than some other numbers, it is because these produce a result where approximately one-half of all points will be attributed to offence, and one-half to defence. I know this to be true, as I have tested it; I just haven't proven it mathematically.
From here we depart from Mr. James' work. The next step is to allocate the OP and DP amongst the team's players, and the methods used in baseball's Win Shares are not transferable to hockey. So I have devised my own methods of doing so. As an illustrative example, I will go through the Point Allocation calculations for the 2000/01 Detroit Red Wings.
Team Analysis
The Red Wings played 82 games, collecting 107 points (note that I have eliminated points for OT losses, to keep consistency with the entire history of the NHL). In a league in which 226 goals was average, they scored 253 goals and allowed 202 goals. Therefore, their MGF was 140, and their MGS was 137, for a TMG of 277. Thus their 107 points are allocated as follows: 54.1 to offence, and 52.9 to defence.
Offence
Fortunately, hockey stats reflect offensive contribution quite well, through goals and assists. We cannot use scoring points, however, because of the arbitrary way in which they combine goals and assists. There is no reason to think that playmaking is 1.7 times as important as goalscoring (which is what scoring points do in modern times, where there are about 1.7 times as many assists as goals). Since there is no way to determine the relative importance of playmaking and goalscoring, we will assume they are equally important.
Therefore, to allocate offensive points, we need to calculate a new stat, Offensive Contribution (OC), which is simply defined as the player's assists divided by the team average assists per goal, plus the player's goals. For instance, Brendan Shanahan had 31 goals and 45 assists in 2000/01, and the Wings had 1.70 assists per goal. Shanahan's OC is therefore 57 (45/1.7 + 31). Doing this for every Red Wing, we find the team total is 509 (which is twice their goals, with a rounding difference). Shanahan's OC is .112 of the team total; therefore he receives 6.1 Offensive Points (OP), which is .112 of the team's points allocated to offence. Note that in this analysis, goals and assists by goaltenders are ignored. A goaltender's value lies in stopping pucks, not in shooting the puck down the ice into an empty net. Similarly, goalie assists are more a function of team offence, and have little to do with the goalie's skill.
Defence
Defence in hockey is made up of two parts: the skaters who attempt to prevent shots, and the goaltender who attempts to stop those shots that are allowed. Therefore, before we can allocate defensive points amongst a team's players, we need to determine how many go to team defence, and how many go to goaltending.
Since we have defined a defence's job as preventing shots, and a goalie's job as stopping shots, we will use team shots against and goalie save percentage in conjunction with marginal goal analysis to allocate points.
We start with team defence. The defence is responsible for preventing shots. Therefore, we calculate the MGS you would expect for the team based on their actual shots allowed. Detroit allowed 2221 shots in 2000/01, and the NHL average scoring percentage was 9.95%. Therefore we would expect Detroit's defence to have a MGS of 118 ((226 x 1.5) - (2221 x .0995).
Now we move on the goaltending. Since a goalie's job is to stop shots, we evaluate them based on their save percentage. We calculate the MGS we would expect for the goalies based on their save percentage, and then calculate a weighted average for all the team's goaltenders, based on the goalies' playing times. Manny Legace had a .920 save percentage in 2000/01. The NHL average shots per game (excluding empty-net shots) was 2265, and the league average goals (excluding empty-netters) was 219. Legace's MGS is therefore 148 ((219 x 1.5) - (2265 x (1 - .920)). Similarly, Chris Osgood's MGS is 109 based on his .903 save percentage. Legace played 2136 minutes, and Osgood played 2834; the weighted-average MGS for Detroit's goaltending is therefore 126.
We now need to combine these two figures. There are, on average, 4.9 skaters and one goaltender on the ice at any one time. Therefore we will assume that the skaters' value is 4.9 times as important as the goaltending value. Therefore, the DP are distributed as follows:
DPS = DP x MGSS /(4.9 x MGSS + MGSG)
DPG = DP x MGSG /(4.9 x MGSS + MGSG)
Where DPS is Defensive Points allocated to skaters, DPG is Defensive Points allocated to goalies, DP is team Defensive Points, MGSS is the MGS value for skaters, and MGSG is the MGS value for goalies.
For Detroit, this works out to 9.5 for goaltenders and 43.4 for skaters.
Allocating DP to Goaltenders
Allocating the goaltending DP among the team's goaltenders is a simple task. Simply take each goaltender's contribution to the team weighted-average MGS to determine the proportion of DPG he receives. For instance, Detroit's MGSG of 126 was made up of 64 from Manny Legace (148 MGS time his proportion of minutes played) and 62 from Chris Osgood. Therefore, Legace receives 50.8% of the DPG (64/126), or 4.8 points. Osgood receives the remaining 4.7 points.
Allocating DP to Skaters
Skaters' defence is such an ephemeral quality; we all know that the stats don't reflect defence in any meaningful way. But wait! Remember what I was discussing before I got into all of this; a player's defensive value is reflected in his playing time, when his offence and his teammates are taken into consideration. Just like Bill James' breakthrough in fielding analysis, we start at the team level. It's probably easiest to explain by diving right into the illustration.
First, we must assume that a skater's job is made up of equal parts offence and defence, on average. Then we also assume that a player's total value to a team is reflected in his playing time; that is, a team's best players will play the most. I believe these are perfectly logical and safe assumptions. We then compare each player to the average offensive numbers for his position (forward and defence) to find his offensive contribution relative to the team average. Comparing this to his actual playing time, we can estimate his defensive value to the team.
Let's look at some numbers. Detroit's forwards, in total, played 14,250 minutes in 957 games, for an average of 14.89 minutes per game. They had a total OC of 386. The average OC for a Detroit forward was therefore 0.40 per 14.89 minutes.
Now let's look at Brendan Shanahan, Detroit's top scorer in terms of points. He played an average of 18.37 minutes per game, and had an OC per 14.89 minutes of 0.57. His OC was 1.425 times that of an average team forward; if playing time depended only upon offence, we would thus expect him to play 21.22 minutes per game (1.425 x 14.89). But he played only 18.37 minutes per game; 2.85 minutes per game less than the offensive expectation. This difference must be due to his defence, which is obviously not as good as his offence. His defensive minutes per game would be 18.37 minus 2.85, or 15.52; since offence and defence are equally important, this will give us his average playing time of 18.37 per game. If playing time were based solely on defence, Shanahan would probably play about 15.52 minutes per game. He played 81 games, so his total defensive minutes would be 1,257 (15.52 times 81).
We do this for each player in turn. Note that for defencemen, the values for the Red Wings defencemen must be used (9,793 minutes in 512 games, 19.13 minutes per game, 0.24 OC per 19.13 minutes). Also note that it is possible that a player's calculated defensive time would be negative (though usually only for a player playing only a few games). Since we are dealing only with marginal contributions, negative values make no sense. Therefore, any negative value is assumed to be zero in all analysis involving Marginal Goals, and throughout the Point Allocation system. Adding up the minutes, we find Detroit's team total to be 25,209 defensive minutes. We use this total to allocate defensive points to skaters, based on their proportionate contribution to the total. Shanahan had 1,257 of the team's 25,209 defensive minutes, or 0.050 of the total. He therefore receives 0.050 of the 43.4 skater defensive points, or 2.2 points. Adding these to his 6.3 offensive points, we find his total is 8.5 points.
A Final Adjustment
We want this method to be applicable across all years for which the data is available. We don't want any distortion from schedule length or roster size to affect the results. Therefore adjustments are included, to normalize the results to an 80-game schedule, and also to 15 minutes per game for forwards and 20 minutes per game for defencemen. For example, Shanahan played 81 of 82 games; we adjust this to an 80-games schedule, so we give Shanahan 79 GP. He played 18.37 minutes when the average was 14.89; we adjust this for an average of 15, so we credit him with 18.51 minutes per game. So his total minutes are now 1,462 (79 times 18.51), instead of the 1,488 minutes he actually had. We then adjust his OP and DP based upon this adjusted minutes value. These adjustments will eliminate any bias when comparing today's players to players from the days when they played 70 games per year, or when only 17 skaters were allowed to dress. Similarly, goalies' minutes are adjusted to a base of 4800.
Here are the complete team results for the 2000/01 Red Wings. GP is adjusted GP, MIN is adjusted minutes, OP is offensive points (adjusted to MIN), DP is defensive points (adjused to MIN), and TPA is Total Points Allocated (the sum of OP and DP).
| Name | Pos | GP | MIN | OP | DP | TPA |
| Lidstrom | D | 80 | 2379 | 5.4 | 3.7 | 9.1 |
| Fedorov | F | 73 | 1550 | 5.9 | 2.9 | 8.8 |
| Shanahan | F | 79 | 1462 | 6.2 | 2.2 | 8.4 |
| Lapointe | F | 80 | 1297 | 4.9 | 1.9 | 6.8 |
| Yzerman | F | 53 | 1187 | 4.1 | 2.5 | 6.6 |
| Kozlov | F | 70 | 1038 | 3.3 | 1.6 | 4.9 |
| Legace | G | 2063 | 4.8 | 4.8 | ||
| Osgood | G | 2737 | 4.7 | 4.7 | ||
| Maltby | F | 70 | 1007 | 1.8 | 2.5 | 4.3 |
| Draper | F | 73 | 987 | 2.0 | 2.2 | 4.2 |
| Gill | D | 66 | 1284 | 0.9 | 3.3 | 4.2 |
| McCarty | F | 70 | 947 | 2.0 | 2.1 | 4.1 |
| Verbeek | F | 65 | 884 | 2.7 | 1.4 | 4.1 |
| Ward | D | 71 | 1261 | 0.8 | 3.3 | 4.1 |
| Holmstrom | F | 71 | 835 | 3.2 | 0.5 | 3.7 |
| Larionov | F | 38 | 699 | 2.1 | 1.5 | 3.6 |
| Murphy | D | 56 | 1112 | 1.4 | 1.9 | 3.3 |
| Dandenault | D | 71 | 1194 | 2.1 | 0.9 | 3.0 |
| Duchesne | D | 53 | 1015 | 1.9 | 1.0 | 2.9 |
| Fischer | D | 54 | 946 | 0.7 | 2.2 | 2.9 |
| Brown | F | 59 | 668 | 1.9 | 0.9 | 2.8 |
| Gilchrist | F | 59 | 693 | 0.7 | 1.9 | 2.6 |
| Devereaux | F | 54 | 550 | 1.0 | 1.1 | 2.1 |
| Chelios | D | 23 | 549 | 0.2 | 1.7 | 1.9 |
| Butsayev | F | 15 | 138 | 0.2 | 0.3 | 0.5 |
| Kuznetsov | D | 24 | 237 | 0.2 | 0.3 | 0.5 |
| Williams | F | 5 | 62 | 0.2 | 0.1 | 0.3 |
| Wallin | D | 1 | 4 | 0.0 | 0.0 | 0.0 |
So here we have objective evidence that Nicklas Lidstrom is, in fact, Detroit's most valuable player. This surprises no one, I imagine. It is worth noting, however, based on the sampling of team calculations I have thus far made, that it is fairly rare for a defenceman to be a team's MVP (i.e., to have the highest TPA). This may seem to indicate that the system has a bias against defencemen. But I'm not sure this is true. A defenceman's job is primarily defence, and is therefore primarily passive. A defender reacts to an opponent's offense. Therefore, he has less control over his defensive contribution than an attacker has over his offensive contribution. This is reflected in the numbers, where DP tend to be flatter in distribution than OP. So while TPA indicates a team's MVP quite clearly, remember that it is not entirely fair to compare forwards and defencemen directly, since their jobs are so different.
Note how the system also provides objective evidence of the defensive prowess of the Maltby-Draper-McCarty line. Each have a DP total greater than their OP, which is fairly rare for a forward.
For comparison's sake, here are the 1975/76 Montreal Canadiens, one of the greatest teams ever iced. The ice times are estimates calculated using my method for estimating ice time.
| Name | Pos | GP | MIN | OP | DP | TPA |
| Lafleur | F | 80 | 1709 | 8.7 | 3.8 | 12.5 |
| Dryden | G | 3580 | 11.8 | 11.8 | ||
| Mahovlich | F | 80 | 1524 | 6.9 | 3.3 | 10.2 |
| Shutt | F | 80 | 1412 | 5.8 | 3.2 | 9.0 |
| Lambert | F | 80 | 1362 | 4.7 | 3.5 | 8.2 |
| Lapointe | D | 77 | 2048 | 4.2 | 3.9 | 8.1 |
| Savard | D | 71 | 1760 | 2.7 | 4.4 | 7.1 |
| Cournoyer | F | 71 | 1097 | 4.8 | 1.7 | 6.5 |
| Risebrough | F | 80 | 1104 | 3.0 | 2.9 | 5.9 |
| Lemaire | F | 61 | 991 | 3.6 | 2.0 | 5.6 |
| Robinson | D | 80 | 1574 | 2.3 | 3.1 | 5.4 |
| Awrey | D | 72 | 1276 | 0.6 | 4.6 | 5.2 |
| Gainey | F | 78 | 1021 | 2.0 | 3.1 | 5.1 |
| Jarvis | F | 80 | 941 | 2.0 | 2.6 | 4.6 |
| Bouchard | D | 66 | 1051 | 0.7 | 3.4 | 4.1 |
| Wilson | F | 59 | 763 | 2.3 | 1.7 | 4.0 |
| Tremblay | F | 71 | 771 | 1.8 | 1.9 | 3.7 |
| Roberts | F | 74 | 783 | 1.6 | 2.1 | 3.7 |
| Larocque | G | 1220 | 3.1 | 3.1 | ||
| Van Boxmeer | D | 46 | 672 | 1.1 | 0.4 | 1.5 |
| Nyrop | D | 19 | 326 | 0.2 | 1.2 | 1.4 |
| Chartraw | D | 16 | 233 | 0.3 | 0.5 | 0.8 |
| Goldup | F | 3 | 21 | 0.0 | 0.1 | 0.1 |
| Shanahan | F | 4 | 25 | 0.0 | 0.1 | 0.1 |
| Andruff | F | 1 | 10 | 0.0 | 0.0 | 0.0 |
No comments:
Post a Comment