#PS2 - Fun with Fenwick, of-Corsi-did

In this second instalment of Hardly a Stat Holiday for the Patrik Stefans, I felt no choice but to examine what has emerged as the preeminent statistic in advanced hockey analyses: Fenwick.  Named after the Calgary Flames blogger, Matt Fenwick, the statistic looks to improve on its predecessor Corsi metric. For those who are not sufficiently acquainted by now, you should be, so here's the simplest explanation.  Corsi initially arose as a team differential value between the sum of its own shots taken, missed, and (those which were) blocked, and that of its opponents.  Let's take this fun example from a game between the Toronto Maple Leafs and the Minnesota Wild from Oct 15, 2013.  The Event Summary from the game gives each team's totals for shots on goal, shots missed, and shots that were blocked (not to be confused with its own totals for shots that its players blocked), and the Corsi differentials can be easily calculated accordingly:

	TOR	MIN
Shots on Goal	14	37
Shots Missed	11	14
Shots Blocked	5	17
Corsi	-38	38
Fenwick	-26	26

Corsi has come to be better expressed as a percentage which takes one team's total "shot events" and divides it by the grand total of shot events for both teams combined.  In the above example, Toronto's "Corsi For Percentage" (CF%) is (14+11+5)/(14+11+5+37+14+17) = 30/98 = 30.61%.  Conversely, Minnesota's CF% is 69.39%.  If we look at this season's NHL rankings by CF%, you'll see just how abysmally low that above figure is for the Leafs from that isolated game. Not so coincidentally, the Leafs are bottom feeding at 44% at the rough-halfway mark of this season.   

#	Team	FF	FA	FF60	FA60	FF%	CF	CA	CF60	CA60	CF%
1	Los Angeles	1639	1306	43.1	34.3	55.6	2175	1686	57.2	44.3	56.3
2	New Jersey	1417	1129	37.8	30.1	55.7	1897	1497	50.6	39.9	55.9
3	Boston	1751	1489	45.2	38.4	54.0	2391	2005	61.7	51.8	54.4
4	Chicago	1565	1286	40.8	33.5	54.9	2103	1783	54.8	46.5	54.1
5	Ottawa	1683	1509	45.0	40.3	52.7	2305	1994	61.6	53.3	53.6
6	Detroit	1563	1368	41.6	36.4	53.3	2069	1793	55.1	47.8	53.6
7	Montreal	1534	1332	42.1	36.5	53.5	2114	1886	58.0	51.7	52.8
8	St. Louis	1455	1267	38.5	33.5	53.4	1949	1784	51.5	47.2	52.2
9	NY Rangers	1674	1456	43.1	37.5	53.5	2253	2081	58.0	53.6	52.0
10	Carolina	1712	1671	45.3	44.2	50.6	2336	2190	61.8	57.9	51.6
11	Vancouver	1457	1438	38.7	38.2	50.3	1994	1879	53.0	49.9	51.5
12	San Jose	1646	1520	42.8	39.5	52.0	2258	2144	58.7	55.8	51.3
13	Phoenix	1646	1597	43.6	42.3	50.8	2173	2117	57.6	56.1	50.6
14	NY Islanders	1591	1539	40.6	39.3	50.8	2172	2174	55.4	55.5	50.0
15	Winnipeg	1597	1644	40.7	41.9	49.3	2195	2241	55.9	57.1	49.5
16	Minnesota	1400	1437	37.0	38.0	49.4	1883	1925	49.8	50.9	49.5
17	Florida	1529	1628	39.2	41.8	48.4	2049	2121	52.6	54.4	49.1
18	Pittsburgh	1528	1533	40.5	40.6	49.9	2040	2120	54.1	56.2	49.0
19	Colorado	1526	1580	40.0	41.5	49.1	2073	2170	54.4	56.9	48.9
20	Dallas	1409	1579	37.8	42.4	47.2	1990	2084	53.5	56.0	48.9
21	Washington	1479	1563	38.8	41.0	48.6	2043	2150	53.5	56.3	48.7
22	Anaheim	1444	1530	37.3	39.6	48.5	1970	2140	51.0	55.3	47.9
23	Philadelphia	1417	1547	38.6	42.2	47.8	1943	2140	52.9	58.3	47.6
24	Calgary	1448	1599	37.7	41.6	47.5	1951	2164	50.8	56.3	47.4
25	Tampa Bay	1451	1625	38.2	42.7	47.2	1921	2151	50.5	56.6	47.2
26	Columbus	1287	1530	33.6	40.0	45.7	1811	2031	47.4	53.1	47.1
27	Nashville	1344	1491	34.6	38.4	47.4	1800	2053	46.3	52.8	46.7
28	Buffalo	1348	1663	36.2	44.6	44.8	1811	2203	48.6	59.1	45.1
29	Edmonton	1379	1680	36.9	44.9	45.1	1816	2262	48.6	60.5	44.5
30	Toronto	1396	1779	36.9	47.0	44.0	1928	2445	50.9	64.6	44.1

What is Corsi essentially trying to measure?

Corsi has commonly been described as a proxy for scoring chances.  That's it.  It's a metric which carries information about discrete scoring opportunities created by a team against scoring opportunities it gives up to its opponents.  We should be careful not to conflate the metric with scoring chances, as that is a statistic which is subjectively recorded separately.  In any case, the proxy for scoring chances is ultimately supposed to give an objective understanding of how proficient a team is at scoring goals vs. giving up goals viz. generating scoring opportunities vs. giving up scoring opportunities; boiled down to its most extreme claim, it is a predictor of winning games.

The aforementioned blogger Matt Fenwick was keen to point out, as hopefully most of you have come close to concluding likewise, that a shot from the point which gets blocked should hardly count towards a proxy value for scoring chances.  Moreover, a shot blocked by one's own team may in fact be more representative of some defensive skill which presumably contributes to the team's success.  Why on earth should it be factored into inflating the opponent's Corsi?  Accordingly, Mr. Fenwick, modified the metric by simply getting rid of the blocked shot category.

Problem Solved?

If anybody is convinced that this newer Fenwick metric is some magic window into success in the National Hockey League, reconsider that conclusion.  I know as well as the next stat hack not to use some token counter-exemplifying instance to rebut or disprove a general relationship, but I can't help but point out that the Maple Leafs actually won the fucking game above against the Wild 4-1, a game in which the respective Fenwick For Percentages (FF%s) were 32.89% and 67.11%! As a strengthening of the case against Fenwick, I can provide with more conviction the correlation coefficient for team pts vs CF%; it is a tepid .4511, meaning that only a very modest relationship exists.  Surely, Brian Burke, a self-proclaimed advanced-stat-hater would be pleased to know this, and that truculence and pugnacity still play an immeasurably vital role to winning.

Individual Fenwick

The Fenwick statistic has been converted in the same way that the Corsi has to measure the individual player.  It's very simple.  "iFenwick" simply gives a player's total shots on goal plus shots missed.  That's all it is.  Again, it is commonly viewed as a a proxy to a player's ability to generate scoring chances.  But again, and as I'm hopeful most of you will have already realized, it does not credit a player for making a great outlet, a clever drop pass, providing a screen in front, making a hit to free up a loose puck in the offensive zone, etc.  All of those latter events often translate into points, yet none are captured by Fenwick.

Fenwick and the KL

In any case, you may believe that it is as good metric out there at predicting player pts, even if it misses out on those key hockey-goal-generating plays, and especially despite of its older cousin's, the team Fenwick's, underwhelming correlation with team points.  Before writing this post and performing the analysis below, I predicted that the iFenwick would in fact be an even worse predictor for player points than the team Fenwick is for team points.  That is because it carries with it the very same pitfalls of the team Fenwick and also fails to capture any information about assist-related events.  It turns out that my hunch was correct.  Below is a table which shows in decreasing order KL teams by their average iFenwick per 60 minutes as of the Xmas break.  Avg iFenwick is actually slightly negatively correlated with KL rank (-0.1152)!!

TEAM	KL RANK	Avg iFenwick/60
Vanrooser Canicks	15	10.04511
Milan Micahleks	2	9.66286
Moilers	8	9.61868
Dicklas Lidstroms	11	9.55842
G-Phil's Flyers	1	9.40037
Fylanders	3	9.37320
Teeyotes	7	9.20821
Quebec Rordiques	14	8.98260
W-Benham/Scranton Parkers	16	8.93172
Patrik Stefans	12	8.81368
Winter Claassics	6	8.72440
Powder Rangers	4	8.69511
Mackhawks	5	8.64050
Joshfrey Krupuls	10	8.53160
Schizzarks	13	8.13147
Los Samjawors Kings	9	7.72316

Discussion and Shortcomings of my Own Methodology

Quickly, before this stat holiday is over in all parts of North America, I will draw attention to a few issues.  First, I used iFenwick/60 instead of straight iFenwick to adjust for players that haven't played many games.  I wanted to give each of you as accurate a value as possible to represent how much your players direct the puck toward the net when they're on the ice, and including, for example, Steven Stamkos' total iFenwick instead of his iFenwick/60 would have been misleading.  That being said, the adjustment so that we are comparing apples with apples (okay, trying hard not to conflate the terms of analogy with hockey slang for assists) might go too far because it's not particularly relevant to the KL how much a player shoots if he's not playing meaningful minutes.  It might be indicative of some potential, especially if management and coaches are big Fewickians and it leads to more ice-time, but that's it, and I would caution against getting too excited if you notice some high iFenwick/60 numbers for some bottom end guys.  I almost picked up Colton Sceviour last week, and I'm glad I didn't.

Another limitation of my analysis is that the numbers only reflect 5-on-5 play.  My suspicion is that this has the greatest impact on defensemen who take many of their shots on the PP.  Not much to say further to that, other than your true iFenwick/60 will differ slightly, and as a rough approximation, you might expect it to go up slightly if you have PP contributing defenders.  The thing is, most of us do anyway.

Lastly, and in addition to reiterating the fact that Fenwick is blind to playmaking, it is also blind to sharp-shooting.  I haven't run the numbers, but I would guess that in addition to holding some good play makers, the rosters above which have a lower iFenwick/60 rank than their overall KL Rank also likely have some really keen snipers.  And by that, I mean, they have some of the most accurate shooters in the game.  Backes, Stamkos, Filppula, Monahan, Foligno, Grabovski, Nielson all fit that description and belong to "underperforming iFenwick/60" teams.

Take from the above what you want.  Many of you already look at this stuff, and some more than others.  To no surpirse, the Michaleks sit squarely in second in both this statistic and the KL standings. However much stock GM Carmody puts into the Fenwick element, it seems to be paying dividends. For others like the The Claassics or 'Jawors Kings, it is perhaps best they continue to presumably ignore such metrics.  

It's back to real work for most all of us tomorrow, but I assure you I look forward to the next Stat Holiday.  NB: You may have noticed that I missed Christmas Day.  This post was initially earmarked for that day, but shortbread and eggnog took priority.  I have another analysis in the hopper that I will publish this weekend as somewhat of a Stat Holiday carry-over.  And after that, I will disappear until Good Friday and Easter Monday, when I'll hit you guys with two more.  That weekend beautifully follows the conclusion of the regular season, so there will be a lot to look back on I'm sure.  Thanks for reading... I know this is fun for at least one of us.