Prediction for July 21-22: Utah 1 v North Carolina 2 Orlando 1 v Seattle 0 Washington 1 v Houston 1 -- I'll be there - first time to Md Soccerplex. Gonna miss seeing Ohai; but I love she is with the Nats.NJNY 0 v Portland 1
Scores are good again for this weekend's games, with lunatica and Smallchief each gaining a brilliant 9 points out of a possible 12. lunatica is threatening CoachJon's long hold at the top of the human rankings. There's only one game for week 18, Chicago v NJNY next Saturday.
Wow, I can't believe I made 8 points this week!!! I wonder if the recent scandal involving Sky Blue's management can influence next results in any way. Despite being in doubt about that, my prediction is : CHI 2 NJ 0
Since we're sort of in a break -- except for Red Stars v Sky Blue -- here's a detailed breakdown of how our predictors are doing, using a "per game predicted" approach, followed by some stats details on the teams as we head into the regular season end game. The following is for the last 98 regular season games played, which means there is a carry over of some games from last year: And this one is for the current season only: Regarding the home wins/ties/losses numbers, some might think they mean that there isn't much home field advantage in the NWSL. But, that's not necessarily the case. Here's how the individual teams have done at home (W/L/T). I've got the teams in the order from the preceding table. North Carolina 6-1-2 Seattle 5-1-2 Portland 4-3-2 Chicago 4-3-3 Olando 3-3-3 Houston 3-2-4 Utah 3-2-4 Washington 2-5-2 NJ 0-8-1 So, North Carolina and Seattle at the top are a combined 11-2-4 (W-L-T) at home; whereas Washington and NJ at the bottom are a combined 2-13-3. The other teams in the middle are a combined 17-13-16. Or, if I put all the teams but Washington and NJ together, they are a combined 28-15-20. Looking at Washington and NJ, to illustrate what I think is happening, this doesn't mean they don't have a home field advantage. What it means is that there's such a big gap between them and the other teams that the gap is too big for home field advantage to make a difference in game outcomes. In other words, it's possible, and I think it's likely the case, that the conglomerated home W/L/T numbers, with about the same number of home wins as losses, are a reflection of a big imbalance between the bottom two teams and all the other teams rather than an indicator of a lack of home field advantage in the NWSL.
It's worth pointing out that looking for HFA is probably better done on a team-by-team basis, comparing their home and away forms. In particular, five teams in the table - NC, POR, CHI, ORL, and NJ - all have better away records than home records when looking at PPG. Even if you assume some combination of NC, DC, and NJ are skewing the home/away comparison and remove their matches from consideration, you still get POR, CHI, and ORL with better away records than home records - ORL and CHI noticeably so. At the same time, other teams improve the HFA benefit by removing results for the suspect teams, so *shrug*.
The problem with the NWSL is, there are not enough data. I've done extensive studies for NCAA Division 1 women's soccer, with a data base of about 35,000 games, and they show that there is a home field advantage there. But, it doesn't affect game results much unless the teams are closely rated since if they aren't closely rated, the better team tends to win regardless of playing away. My hypothesis is that the same is true for the NWSL. But, there aren't close to enough data to show that. PS -- I wasn't comfortable establishing the extent of HFA for DI women's soccer until I reached the 30,000 game threshold. So, from my perspective, any attempt to define NWSL HFA is premature to the extreme. Beware the law of small numbers that human beings tend towards, but that causes all sorts of problems (which law is: Any data sample is good, no matter how small).
I think saying "not enough data" depends entirely on what confidence level you're talking about, and what assumptions you go in with. From what I've read of professional statisticians (e.g. political polling, big data crunching in scientific articles, sabermetrics, etc.) you can get to a margin of error of about 3% with only ~1000 data points*, and within 5% using a hundred or two. We're still an order of magnitude less than that for NWSL, but even that depends on exactly what question you're asking, e.g. are you talking about one season for one team in particular, or the league in general over its entire history and all teams. That's the difference between considering only a dozen or so games versus considering several hundred. Either way, having 30k+ data points is probably overkill unless you're using rather stringent filters, such as what I think you're doing by only considering very narrow bands of similar-strength teams. HFA is particularly an interesting issue since there are several ways and degrees to which is can be defined. Your signal-to-noise ratio is gonna be different, even with the same dataset and same other assumptions, if your define HFA to be as narrow as winning 51% of non-drawn games versus winning 60%. Soccer also has complicating factors such as the size and material of different fields or the amount of noise home fans produce. In theory, each game can thus be weighted differently as to whether you think it'd be better or worse at proving HFA, which could help or hinder your analysis depending on the question you're asking and filters you're using. FWIW, in the area of study I'm getting my degree in, most analyses are done on datasets that are on the order of 10-50 data points. Sometimes the signal-to-noise ratio is low, but it's almost always enough data to at least prove some patterns and correlations exists - especially so for some cases when you have more information about the physics in the experiments that lead to each data point. So that's the origin of my perspective. Incidentally, 20-40 data points is often the number cited for a lot of general pattern identification issues - establishing someone's rating in an Elo system, going through Phase 1 of clinical research, etc. Enough to confidently identify and qualitatively describe a pattern, though maybe not enough to quantitatively nail it down to rigorous standards. (*that ~1000 data points thing mentioned earlier, for example, is not only the accept level for professional polling, but also the number of people usually involved in Phase 3 of clinical research, the final step before drugs or procedures get cleared for widespread use). === ...this is all an exercise in semantics at the moment, since I feel that applying enough filters to suggest HFA is present within just this season does get us a bit too close to the "small numbers" limit, which is why I'm much more comfortable at the moment to say that HFA this season is negligible, if present at all. You also have to apply the filters correctly... If you're using a strength filter, you have to apply it at the top as well as the bottom, and grouping SEA with NC isn't the proper grouping based on performance. So that 28W/20L split should be 22W/19L if you're looking at the 6 teams that have been grouped together for most of the season. Also, if you discount just the two bottom teams but not the top team, the same analysis on home games that gives the 28W/20L "home advantage" would also produce an "away advantage" of 28W/18L if you look at just away games; either way, you've simply eliminate losing records from an otherwise roughly even split, so I would say that doing an analysis that looks at just home games or just away games but not both is applying one filter too many.
STT, great post and very interesting. I'm most interested in quantifying average home field advantage for D1 women's soccer using, as a base rating, the NCAA's rating system (and I've done it for an Elo-based system, too, with comparable results). What I determine is the rating value of being the home team, to determine a rating adjustment to apply when two teams are playing each other and, with the ratings so adjusted, the win/loss/tie likelihood for the game. To get what I believe are the most accurate results for the "rating" value of HFA, I look at the closest 5%, 10%, and 15% of games. Thus you're right that I'm using stringent filters. In fact, I'd prefer to use the closest 5% but don't think I have enough data yet based on the variability in the 5% numbers I see each time I add a new year's data. (Bottom line, I'm interested in the maximum possible precision with minimal margins of error.) Interestingly, for D1 women's soccer, the best rating system (Massey) computes a separate home field advantage for each team (whereas I use an average HFA across all teams). Based on the variations I've seen in HFA with the vastly greater numbers I'm using, I'm skeptical about the small numbers he must be using, but ... he has the best performing rating system (meaning that his ratings correlate the best with game results). What I've seen is that HFA is not a large factor. It's there, but it can be overcome by stronger teams. The 22-19 split for the NWSL teams that have been grouped together fits my sense of HFA. And, I just realized that I have individual team data on HFA: This table (using the last 98 games, which includes the later games from 2017) doesn't take into account which teams were the away opponents. With that qualification, it suggests that all of the teams but North Carolina (which posters have noticed) and Orlando have had a home field advantage. But this is where I think there may be insufficient data. I think that North Carolina and Orlando also may have home field advantage and, given an infinite number of games, might show up on a table like this with HFA too. At the same time, for those teams with a table-demonstrated home field advantage, the advantage for most teams isn't particularly great and probably isn't enough to determine the outcome of very many games.
Just for fun as to the value of home field advantage, I went to my Division 1 women's soccer data, looked at the average value of HFA there, and came up with an equivalent HFA for the NWSL. If HFA is comparable between the two, then it's worth roughly 1.56 NWSL ranking points. Or, to put it differently after rounding up to 2 ranking points: the home team plays as though it has 1 more ranking point than it has and the away team plays as though it has 1 fewer ranking points. So, using Seattle v Portland as an example since they've both played 18 games, Seattle has 30 ranking points and Portland has 29. For illustration, let's assume they've played the same opponents. Then at a neutral site, Seattle would be a very slight favorite over Portland. For a game at Seattle, it would be as though Seattle has 31 points and Portland 28, so Seattle would be slightly more of a favorite but still only a slight favorite. For a game at Portland, it would be as though Portland has 30 points and Seattle 29, so Portland would be a very slight favorite. The point being, home field advantage, on average, is quite small.
Should NJ have an away win by 0-2 instead, we'd basically all bankrupt! Edit: by the way, are we all taking into account that CRS will be missing Kerr because of ToN?
I actually believe that Sky Blue is a little better than its record indicates. And Chicago is missing Kerr (I think) and a few others who are on the NT.... I think Sky Blue is missing Katie Johnson, on duty with the Mexican team, who may be their best forward.. But, what's her name? McSomebody, is looking better all the time as a center forward or 10.
One game. Play it safe and copy the robot? Or, look back at Lunatica, who is charging like Arnold Palmer? Or look at BlueCrimson, who has been always right there? None of the above. CHI 1 v. NJ 0