The statistical aficionados at fivethirtyeight.com have added Major League Soccer to their regularly updated tables of league soccer predictions. Seattle Sounders edge out Toronto FC and New York Red Bulls in predicted points and chance at winning MLS Cup. As is the case with so many things 538, the prediction table is a statistical model: it doesn’t incorporate any of that secret insider knowledge you’ve used to fill out your own personal power rankings or fantasy team, so there’s no point in quibbling with the authors on those points (let alone me). That objectivity can be both a strength and a weakness, so here’s a few things to keep in mind about the methodology for this table.
The 538 predictions are powered by a “Soccer Power Index” ratings system 538 founder Nate Silver originally developed for ESPN some 7-8 years ago. The system uses a combination of goals scored and conceded along with analysis of expected goals from shots and potential shooting situations to generate defensive and offensive ratings for each team in a league (the values translate roughly to goals against a generic opponent, so a high rating is good for offense and bad for defense).
They then runs a Monte Carlo simulation of the season with 10,000 iterations. Head-to-head results are determined from a probability distribution of results based on each team’s ratings, home field advantage, and team rest in the schedule. Based on the simulated result, each team’s ratings change dynamically over the course of a simulated season (note, interestingly, that this method could subtly lead to encouraging streaks). On the prediction table, you can select other leagues around the world. Perhaps the most glamorous example of SPI’s application (and one you can use, anecdotally, to judge the application of the model to real-world competitions) is the 2014 World Cup (actual results, for comparison). Seattle averages 54 points over the season in the current simulations. The table will be updated as results come in.
A few other things to keep in mind:
- Objectivity - All teams are processed by the same goals-based method that serves as a more effective predictor of success than a team’s won-loss record.
- Resistance to luck - Luck, both in that a team may have good fortune in results from a given goal differential, and also in a team’s good fortune of having a particular shot turn into a goal. 538 averages 4 goal metrics: actual goals, goals weighted by situation (garbage time padding of a lead, for example, is down-weighted), expected goals from shots, and expected scoring situations (say, an intercepted ball in the 6-yard box).
- Goals may be a poor measure of quality. Statistical metrics to determine which team will win the midfield battle and, therefore, unexpectedly silence an otherwise lethal offense don’t really exist. 538 acknowledges that goals happen infrequently and soccer results are therefore subject to a large factor of chance. That’s just part of the sport.
- It’s a challenge to gauge offseason change in a league like MLS. The SPI metric in its ideal form takes into account player ratings alongside team ratings, though I’m not sure exactly the extent to which this has been done for the MLS ratings, where possible (e.g., the player ratings are dependent upon the league having full stats via OPTA, so Minnesota’s NASL players are outta luck). For example, the addition of Minnesota and Atlanta to the league seems to have been addressed through some kind of brute force (their ratings are identical), and the change to the league’s competitive balance will likely have a trickle-down effect throughout the table. No statistically objective model deals well with missing data.
- Parity. The competitive difference between the top and bottom of the table in MLS is less than that for many other leagues. Small changes may cause a great deal of volatility in the table (consider the limited moves that turned Seattle from cellar-dweller to champion).
- Conservatism. Whether Seattle (and other teams, for that matter) would make the playoffs only 8 of 10 times is a matter of debate, but expecting Chicago in the postseason 4 of 10 times reflects both the aforementioned parity and a certain fuzziness in the model. The simulated range of goal difference and record difference in the league is likely less than will turn out in reality.