Statistical Projection

Introducing Playoff Seeding Projections by Drew Olsen

As you've probably seen under the Projections tab in the upper right, for the last couple months we've been keeping the odds for each MLS team's chances at making the playoffs and winning the Supporters' Shield. 

Our playoff probabilities come from a combination of 1) where teams are now in the tables, 2) what their remaining schedule is, and 3) how good our model thinks they are. It's the same model that produces the Power Rankings, but the key difference is that here we take each team's current standing and remaining schedule into account.

Read More

Our Playoff Chances Model is underselling FC Dallas by Matthias Kullowatz

By Matthias Kullowatz (@MattyAnselmo)

Our on-site playoff chances finally gave FC Dallas a better-than-50-percent shot to make the playoffs after its away win over Chivas on Sunday. I say, "finally," because despite being in a playoff position much of the season, that particular playoff model hasn't been too convinced. 

It's important to understand how our playoff model works, and what its weaknesses are. It is based on overall shot ratios and finishing ratios, though at this point in the season it's the shot ratios that dominate the model's predictions. If you take a look at our MLS Tables, you'll see why the model isn't too keen on FC Dallas--its shot ratio of 0.84 shows that it only tallies 84% of the number of shots that it gives up to opponents. So despite being in fourth place by points per game, that model likes Vancouver and Colorado more because of their respective 1.03 and 1.30 shot ratios.

I don't yet have enough granular shot data to form a full playoff projection model using our Expected Goals 2.0, but we can still use that to intuitively tweak our expectations. Here's why FC Dallas is better than its 0.84 shot ratio.

Expected Goal Differentials (xGD) are more predictive of future success than simple shot ratios. At least, the 1.5 seasons of data we have say so. xGD takes into account not just shot quantity but also shot quality, based on the shot's origin and which body part was used. Based on 2013 and 2014 data--when controlled for number of home games--just eight games of xGD information predicts the following eight games of actual goal differential with a linear correlation coefficient of 0.33. Bump that up to 17 games of xGD information, and one could predict the following 17 games of actual goal differential with a linear correlation coefficient of 0.89 (based on 2013 data only). xGD is strong stuff.

FC Dallas' current xGD of +0.05 ranks fourth in the Western Conference, better than its seventh-place ranking in shot ratios. That's good news for Hoops fans, but I can get an even better idea of how they'll play if I break down xGD further into some specific gamestates.

Watching your favorite MLS team on the road while it's tied or ahead is a nerve-wracking experience. Away teams sitting on points often hang back and allow the home team to suffocate them, hoping to bend but not break. These scenarios are not exactly indicative of the away team's ability. But home teams usually play how they want to, regardless of gamestate, and thus all of a team's past Expected Goals data from when it was at home is helpful for projection.

Based on 2013 and 2014 data, the two best Expected Goals statistics to use when projecting game winners 20 weeks into the season are the home team's past home xGD and the away team's past away xGD in -1 gamestates. The data tells me that the best time to really see an away team's ability is when it finds itself behind by a goal. Interestingly, most teams have played fewer than 300 minutes while down a goal on the road--only about three game's worth--and yet that data in combination with home xGD is more predictive than overall xGD. So far, anyway.

As an example of how these two distinct models think of FC Dallas, we need look no further than its next two games---home against Colorado and away against San Jose. The playoff model we use (based only on shot totals, remember) has them with 38% and 21% chances of winning each game, respectively, and an expected point total of 2.3. Compare that to the Expected Goals model utilizing scenario-specific data, which projects them to win with 51% and 58% probabilities, and an expected point total of 3.7.

It turns out, FC Dallas ranks pretty well in both the aforementioned categories, but it should be noted that FC Dallas has played the third-fewest minutes while trailing by one goal on the road. So its variance in that department is greater than that of the typical team. But while the model's estimates as well as the team's outputs are subject to a modest margin of error, there is little doubt these are important gamestates. I leave you with a sortable table for home performance (xGDhome) and away performance when down a goal (xGDaway(-1)).

Team xGDhome Minutes xGDaway(-1) Minutes
SEA 0.93 957 0.17 249
SKC 0.90 1056 1.45 115
LA 0.89 955 0.10 123
COL 0.73 1047 -0.24 195
NE 0.61 963 -0.52 315
TOR 0.61 963 0.92 141
VAN 0.52 966 -1.12 94
FCD 0.46 1067 1.83 103
NYRB 0.43 966 0.00 434
PHI 0.41 863 0.74 326
RSL 0.31 1065 0.02 126
CHI 0.22 1061 -0.92 78
HOU 0.12 964 -0.88 256
SJ 0.10 1150 -0.34 271
CLB 0.05 953 0.89 225
POR 0.04 1074 -0.35 162
DCU 0.02 1153 0.47 124
MTL -0.13 1054 0.02 298
CHV -0.36 868 0.51 162

Expected Goal Differentials are per 96 minutes of play.