2016 U.S. Women's Soccer- Blame it on the Rio Olympics- Should Have This Team Been There? Part 3- What a SNAFU!
2016 U.S. Women’s Soccer- Blame it on the Rio Olympics- Should Have This Team Been There? Part 3- What a SNAFU! A look at the USWNT and NWSL statistically.
The previous articles from this series, the “Blame it on the Rio Olympics” Parts 1 and 2, look at evaluating the U.S. Women’s National Team (USWNT) players statistically to see if the best players from the NWSL were taken to the 2016 Rio Olympics. In Part 3, players are analyzed once again with a different, but now much improved method. In this review we will be statistically evaluating players by “SNAFU”.
Now, this is NOT the old World War II phrase- “Situation Normal- All Fouled Up”. Nope, this is “Statistically Normalized- All For ‘U'”! Actually, that is such a horrible and corny name, so how about “Statistically Normalized – All ‘Fudged’ Up”? After all, I do use a few “fudge factors” to make the statistical math work. Again, this is not an “official name”, it is just a quick name for convenience.
I wanted to be able to use a statistical method that can be used to evaluate NWSL players over the course of a season. But I also wanted to be able to compare teams in a head to head matchup. That is what I hope to accomplish with “SNAFU”. In the next few days I will be reviewing the Sweden versus USA in the 2016 Olympics Quarterfinal. But here, we will use “SNAFU” to evaluate NWSL players to see which players “deserved” to go the Rio Olympics.
Stat Geek Warning! This article contains over 5000 words and a few really bad jokes. This article might take about 15 -20 minutes to read.
In case you want to see the prior statistical analyses.
But here is a video in case you are bored already.
Now that I have probably “jumped the shark” (^^^) in this episode, let us begin…
Introduction
I took the statistics from roughly the first half of the 2016 NWSL season from a site called WoSo Stats and also from the NWSL league website (links are in the ‘Sources’ below). There are 10 Categories and each is worth approximately 10 points. I set the average of each category to about 5, hence my loose definition of “normalized”. Therefore, a score of 50 points means that you are average. A perfect score of 100 means that you are the greatest player ever. A score of 40.25 in this index means that you at least showed up to practice and suffered all of the games on the bench. A score in the 30’s means that you definitely should have stayed on the bench. The categories all have different ways of calculation.
Preliminary Statistical Overview
This was a “quick and dirty” calculation, no official statistical processing was done. By rights, a good statistical process should have something like a T-test performed to see if your data is normal. And that is just the first step before you can do anything fancy like a correlation or regression analysis in order to make accurate calculations with the data. I am not a professional statistician, but I wanted a calculation that can evaluate players over a full season or compare game to game.
“Should have there been ‘hard-core’ statistical processing done”?
The answer actually is NO. What we have for data is half of the 2016 NWSL season or about 50 games. So, to test what statistics there are, I tested a couple obvious goal scoring opportunities, “through balls” and “crosses” that were completed, not overall attempts. If you get low correlations and regression coefficients with these categories, then essentially, any other correlation for “possession skills” or “defensive skills” will be a zero correlation, and a complete waste of time.
So, here is my quick data table:
| Category 1 | Category 2 | Correlation coefficient | Regression R-squared coeff. |
| Through Balls | Goals | 0.2307 | 0.0532 |
| Through Balls | Assists | 0.4201 | 0.1765 |
| Crosses | Goals | 0.2302 | 0.0530 |
| Crosses | Assists | 0.4671 | 0.2182 |
I also did a quick correlation comparing assists and corner kicks and got a 0.48 correlation. Unfortunately, my limited statistical knowledge tells me that these are weak correlations and terrible regression R-squared coefficients. Any use of the regression slope and Y-intercept in an effort to come up with a fancy statistical calculation similar to what was done with the McHale, Scarf, and Folker would be a huge investment in wasted time. The reason why there are weak correlations above is simple, there are lots of ways to score goals! Heck, even the other team can score goals for you with an “own goal”. But notice that the correlations are much stronger for assists. Therefore, it is likely that these were actual assists that led to the higher correlation coefficients. The problem is not everything leads directly to an assist either.
Therefore, it would be far more productive to try to come up with a “logical analysis”. In other words, the statistics should have a meaning and purpose. The system albeit will not be “perfect”, but the overall purpose is for comparing player skills and overall team performance. You want to see some differences, but then again, you do not want one factor dominating to skew the results of players with only one skill set.
“SNAFU” Introduction
The overall theory behind the numbers is that if you help with a goal then you should receive a point. Therefore, an assist is worth a point. Crosses and corner kicks that are completed are almost a point as you are giving your team a chance to score. A shot on goal is also worth a point. A goal is now worth 3 points. 2 points for the goal itself, and a point for the shot on goal. This may not be intuitive at first, but this system can be looked at to evaluate overall team performance. I will describe team comparisons with more detail in the “Conclusions”
So, a “goal point” overall would include all of the dribbles, defensive work, and attacking passes that eventually led to the goal. So, in theory, a team averaging 51 points should have a goal advantage over a team that has players averaging 50 points if they were to play using these statistics.
“Possession fudge factor” explained.
Many of the parameters below deal with “possession”. According to my math, about 15 possessions are equal to a goal (or about 6.7% of the time). I looked at total goals in the league and did a simple comparison to the categories listed. There were about 15 “possessions or defensive stops” for every goal. For example, there were about 16 “tackles” for every goal, and about 15 times per goal for every “dribbled” . The number “15 or 16 times” for every goal did not always apply to all parameters. However, there was enough evidence to indicate that a possession change every 15 or 16 times might lead to a goal or at least a chance at goal.
So from above, using the theory if an assist is now doubled to become a point, then once out of about 7.5 possessions should have a “chance” at goal (or roughly 13.4%). The number 0.134 now becomes my “default number” for most statistics dealing with “possession”.
So, if you lose the ball to a defender you are deducted 0.134 points. Steal the ball right back and you gain that 0.134 points back. There are some exceptions to this, but overall, this is a “rule of thumb” for the “SNAFU possession calculation”.
Categorical Fudge Factor explained.
If you notice in the Category descriptions below, most categories are sums that are adding to “4.5”. Why this number? The reasoning is simple. I wanted the overall average of the categories to be about 5 points each. Therefore, if you “do nothing” all game long, in theory, you are hurting your team. So, with a minimal amount of effort in each category, then you are at least close to average,
See the below categories descriptions for how each category, and all of the parameters are calculated. Keep in mind that the below calculations were set to “per 90 minutes” for each parameter to get a season average for each player. This means now you can use the same exact calculations for comparing teams in a single game as the overall results would be the same.
.
Here are the 10 categories in no particular order at all.
1. Appearance
This has nothing to do with “looks”. This is the amount of time played on the pitch. Reward players who play.
Note:
When comparing teams either “head to head” in one game or in a season long comparison, this statistic is “dropped”. When comparing averages of teams, “bench players” tend to dilute the overall results with this statistic. “Appearance ” is only good for comparing players as the amount of time played should be a factor. Obviously, you do not want a player with minimal playing time to be rated “number 1” if they only played 15 minutes that season.
2. “Personal Team Suffering or Punishment”
I do not have a great name for this. Yes, this is how much “pain” you “inflict” on your team. Reckless fouls, careless yellow cards, red cards, and own goals all cause pain to your team. I looked at several parameters to see ways that a player can hurt or punish their team.
Parameters and their math is below:
Overall Calculation:
Note:
Therefore, if you are really bad, in theory, you can get a negative score here. A non-National team player listed on both the McHale et al. and the “KISS” lists fell about 70 places in the overall ranks due to this category alone. Obviously, she did not make the recommended player list this time. With this calculation alone, “SNAFU” already has an advantage over the prior 2 statistical systems.
3. Goalkeeping Rank
Not much effort was put here, but I wanted a simple but effective way to distinguish the goalkeepers. I split 2 categories worth 5 points each.
Note:
The goalkeeping results of “SNAFU” are actually encouraging. Abby Smith of Boston had the highest goalkeeper rating of 8.88 points. But Hope Solo did well at 8.56 points, Ashlyn Harris was at 8.44. Alyssa Naeher was at 8.28. An average goalie would get 7.42 points for the league average of 1.33 goals against average, and the 75% save percent. So, the goalie stat worked okay by itself. However, the goalie average was not near 5 as I wanted.
4. Bonus
“Why 9 and 1? Why not other numbers?”
Team comparisons
Please note that when comparing teams, then you can eliminate this statistic along with the “appearance” category. All the players receive the same bonus, therefore, it can be eliminated to help see the differences between teams.
5. Possession
4 parameters were measured. “Net Aerial Duels Won/Lost” , “Net Take Ons Won/Lost”, Recoveries, and Dispossessions (Total).
5 parameters were studied here.
Note:
Also, some of the articles listed in the Sources section below give some statistics on goal success rate using some of these parameters. The numbers I have are in the range of some of these sources.
7. Passing completions overall
Notes:
The factor of 0.067 was used as this is the “original possession number”. The 0.134 factor implies that you are “gaining or losing possession”. The 0.067 number implies that you are “keeping possession with minimal effort”. “Ordinary Passes” do have a 75% success rate, and should not be rated as highly than other parameters that have a higher degree of difficulty.
I have always felt that if you keep passing the ball to the wrong team, they will eventually score on you, and that would show up in this category.
8. Set Pieces
3 parameters studied were free kicks, corner kicks, and throw ins.
Source Notes. For more information concerning set pieces, crosses, corner kicks, etc. see the below: I did not use the data from these per se, but used the information as a rough guide.
9. Goal Scoring Rating
Overall, add up goals, assists, shots on goal per game, and minus the “missed shots”.
6 parameters were studied here:
Add them together (and subtract the “dribbled”), and multiply by 0.134. Then add the 4.5 fudge factor.
Another video to give you a break from reading…
The results were added up and then placed in a hermetically sealed envelope. No one has seen the results until now. ???? As a reminder, here were the players invited to the Rio Olympics in 2016. The number on the left is their total score when all 10 categories were added up. The number on the right is their overall rank of all NWSL players in the first half of 2016.
Goalkeepers
Defenders
Midfielders
Forwards
Alternates
Mallory Pugh, Carli Lloyd, and Megan Rapinoe scores did not change overall, and their reviews are in Part 1.
These USWNT players should have been invited to the Rio Olympics.
Goalkeepers-
Defenders
Midfielders/ Forwards
These USWNT Players are maybe’s
Defenders
Midfielders/ Forwards
Maybe they should have stayed home?
NWSL Players that should have been considered!
Defenders
Midfielders/Forwards
Best Player in NWSL from the Wrong Country
Kim Little is again at the top of the list at number 2 with 59.8 points. Maybe we can make Scotland the 52nd state of the U.S. after Canada becomes the 51st? District of Columbia can wait a few more years! I would invite the United Kingdom, but they would just “Brexit” us anyway. (I will apologize to all Canadians and Scots and British people for that terrible joke, but we need all the help we can get in the U.S – and not just in soccer. ????
Conclusions
McHale et al. favors the attacking player, whereas “KISS” favored the pure defender. With SNAFU, there is an advantage of having more distinct categories to reward players in various ways. Therefore, with this system, the best players are the most versatile and have more complete overall skills This is intuitive in a way. In theory, you can have a great defender with poor passing skills. But if this defender keeps passing to the wrong team, the team will lose eventually. So, this theoretical defender might not show up on the McHale et al. list because they do not score. But might be top 20 on the “KISS” list. But with “SNAFU”, the 2 categories would offset each other, and the player may show up as average.
The McHale et al. system was flawed as they were only able to use statistics with low “p-values”. This is, of course, the proper way to run a statistical analysis. However, this limited the amount of viable statistics to determine a player’s net worth. From the simple correlation and regression data above, we should have seen values close to 75% correlation coefficients for the parameters of through balls and crosses as you are attacking near the penalty box. Also note that I used the completed passes on the correlation study. A weak correlation with an obvious goal scoring opportunity means that there is no way that anyone can use “sound statistics” to determine a player’s net worth. The reason why is because defenders who do not attack, such as Becky Sauerbrunn, are vastly underrated in McHale et al. system.
The “KISS” system also has its flaws. The benefit was obviously simplicity, but with its “simpleness”, comes a downside. With “KISS”, the players were mostly ranked in various categories. This is fine for a season-long comparison of players. But, obviously, you cannot use a ranking system for a game comparison between two teams.
Therefore, a system like “SNAFU” is perfect. On one hand, if you want to do a season end comparison of the players, then you change all of the statistics to an “average per 90 minutes”. This gives the player her “typical game” of the season, and would eliminate the unusually good or poor performances that she had.
And for comparing with a game data between two teams, one of the adjustments is that the data is calculated “as is”, and you do not adjust everyone’s statistics to “per 90 minutes”. This inflates the “bench player” in a game case. Imagine if they played for 1 minute and scored a goal, then the “goals per 90 minutes” would be 90!!!
The other adjustments were mentioned before, as there are no “appearance points” and no “bonus points “. So, when you are done, you look at the team average of the players in the game, then the difference in average team scores should show the difference in number of goals of one team over another, at least in theory.
Look for my review of the actual game statistics between Sweden and the USA at the 2016 Rio Olympics. I will have “SNAFU” player analysis to determine which players were the best statistically in the game, and to see if players may hurt their team with their performance.
In the same article I will also briefly post data from a “friendly” played in October between the USA and Switzerland. This will introduce the testing done to see if my theories with “SNAFU” actually work or not.
Will this statistic work on one game versus a half season worth of data? We shall find out!
Another video before sources and acknowledgements:
Acknowledgements:
I want to thank Alfredo Martinez Jr. and everyone at WoSo Stats for making the data in this series possible.
“Thank you” goes to the official statisticians at the NWSL league website.
Also, thanks to McHale, Scarf, and Folker for their published paper in which was the basis for Part 1 of this series.
Sources:
I also placed some articles about soccer statistics in the list below. I may or not have referred to these above.
More from MLS Multiplex
This article originally appeared on 
