Playoff/Shield Push 2014

 Eastern POFF% Chg% Shield% TOR 77.7% 8.3% 4.5% DCU 82.2% 10.5% 4.1% NE 84.4% -5.2% 5.2% SKC 87.0% 5.2% 5.7% CLB 40.3% -4.4% 0.3% NYRB 36.8% 8.7% 0.3% HOU 17.2% -5.0% 0.0% CHI 26.9% -13.1% 0.1% PHI 21.3% -0.7% 0.0% MTL 4.4% -2.6% 0.0% Total 478.2% 0.0% 20.3% Western POFF% Chg% Shield% SEA 98.2% 2.2% 42.9% RSL 72.0% -9.2% 3.4% VAN 80.3% -6.3% 6.9% COL 70.3% -11.5% 3.2% LA 85.8% -3.3% 11.8% FCD 25.5% 11.2% 0.2% POR 27.1% 14.0% 0.2% SJ 20.8% -2.6% 0.2% CHV 2.5% 1.2% 0.0% Total 482.3% 0.0% 68.7%

POFF% refers to the chances of making the playoffs. Chg% tracks the change in each team’s playoff chances from last week. Shield% refers to the chances of finishing with the most points overall.

1) How does simulation work?

I calculate the probability of the possible outcomes (home win, tie, away win) in the remaining games based on each team’s shot data so far this season. Then, I let the simulator go through and pick which games are won and tied based on the those probabilities. If SKC has a 70% chance to beat NE at home, then the simulator will pick them 7 of 10 times, regardless of other simulated outcomes. All game outcomes are chosen in that manner, and at the end of each simulated season, the computer adds up all the teams’ point totals and then ranks them within their respective conferences. At this point, some teams have made the “computer playoffs,” and some haven’t. After 10,000 such simulations, we get a good idea of the proportion of times each team made the playoffs or won the Supporters’ Shield.

Follow-up: Does your crush on Graham Zusi influence the model?

Good question. No.

2a) Sooo…you only use data from this season to predict?

Yes, for now. Last season’s influence on this season is noisy, I’ve learned.

2b) Which data? Does this model use your fancy Expected Goals 2.0?

No. I would need 2011 and 2012’s shot-by-shot data to create a viable prediction model using Expected Goals. That’s for someday in the future. These predictions are calculated based on shot attempts, finishing rates, and to some extent, strength of schedule.

3) Why don’t the probabilities add up to something nice and round?

This model is not cool enough to account for tie-breakers. Thus, these are the probabilities not including ties in the standings for 5th in a conference or 1st overall.

4) How could a team like Portland get a big win on the road, and only gain 2.0% probability of making the playoffs? (which happened a few weeks ago)

It’s a good question. After all, Colorado won at home against Montreal as we should have expected, but gained 10.4%, a lot more than Portland. Because the model is based on shot attempt differentials, a team can improve it’s chances in two ways: by earning more points than expected in a match and/or by significantly improving its shot differentials. Colorado outshot Montreal 19 to 7, increasing it’s shot attempt differential by nearly a full shot per game. Though Portland got a surprising three points on the road in New York, the Timbers were also outshot 15 to 10, and they lost about 0.2 shots of differential on average.

3 thoughts on “Playoff/Shield Push 2014”

1. silvernic says:

Are you comfortable with a model that only gives RSL a 42% chance of making the playoffs?

• Matthias Kullowatz says:

Haha, no, I am not. I am re-running the simulation as we speak because it was too mean to teams performing exceptionally poorly and too nice to teams performing exceptionally well. It was originally based on model coefficients derived from weeks 6 – 15 of past seasons, but it turns out we’re only just coming up on week 8. My bad. The new model will be based on weeks 4 – 10 of past seasons, which should be much less biased in regards to extreme teams.

• Matthias Kullowatz says:

The updated version still does not like RSL that much. I wrote a few articles last week about how RSL could be model-busters, but that it’s a little too early to tell. This model specifically hates the fact that they give up 17.3 shots per game while only getting 11.1 shots for themselves.

I personally think the truth lies in the middle. They are better than the models, but not as good as their current points/game suggest. Someday our model will be able to take more inputs, and hopefully account for more information about each team.