For 2014, we will again be producing MLS tables with many of the statistics we featured last season. The incredibly awesome new addition will be Expected Goals 2.0---for both teams and players---based on our new formula. I explained Expected Goals 1.0 already, but we can use the shot location (based on the above six zones) to calculate expected goals. To view the expected goals data for this season and the 2013 season, simply hover over the Expected Goals 2.0 link along the top bar and the options will drop down.
The primary improvements over xG 1.0 include the following:
- Adjustments for headed vs. kicked shots
- Adjustments for patterns of play
- Expected goals during only even game states
The data clearly say that headed shots are harder to convert than kicked shots from within each zone. This should come as no surprise to soccer fans. Teams and players are also able to replicate their headed opportunities to some degree, so it's only logical to adjust a shot's value for whether it was headed or kicked.
In terms of each shot's pattern of play, Expected Goals 2.0 is calculated slightly differently for teams and players, respectively. Our objective for teams is to produce metrics (for and against) that suggest how a team should have performed, and thus how they are likely to perform in the future. So how a team finished its penalty kick opportunities is not as important since teams are unable to replicate their number of PK chances from the first half of the season to the second half. However, for players, our objective is to figure out how they finish in all situations, regardless of whether or not they will get those same chances in the future. Accounting for an individual player's PK chances, then, is very important.
The two methods do not differ all that much. The Expected Goals 2.0 for both teams and players includes adjustments for the six shot locations and for headed shots versus kicked shots. However, where players' expected goals are adjusted for all patterns of play,* teams' expected goals metrics are only adjusted for the shots they've earned off corner kicks, throw ins, and free kicks. The other patterns of play are lumped into one category since teams are unable to replicate opportunities from one half of a season to the next.
We will also be including a team's expected goals for and against during even game states (i.e., when the score is tied). Predictive studies revolving around the even-game-state expected goals metrics will be published, as well. Since teams often alter their strategies when they take a lead or fall behind, it seems like future predictions could be improved upon by dissecting the data a bit further in this respect.
The Nitty Gritty
A logistic (binomial) regression was utilized to estimate finishing rates for shots from each of the various combinations of shot location, body part, and pattern of play. Perhaps when the data set grows large enough, a non-parametric method can be applied, but until then we believe the best estimations will come from a parametric regression.
*Patterns of play include Opta's following designations: regular play, free kick (direct shot), set piece (indirect), corner kick, throw in, fastbreak, and penalty kick.
League Shot Data 2013 - 2014
Dist refers to the distribution, or proportion of events that occurred from a particular shot location.
Accuracy refers to the proportion of all shots that were on target.
FinishPct refers to the proportion of all shots that were scored.