SEBA Projections

Updated MLS, USL, and NWSL SEBA projections through September 24

Below are the updated season forecasts using data from games through September 24.

Power Rankings

The “Power Rankings” we concoct are the “strength” of the team according to its competitive expectations. They are computed by forecasting the expected points (3 x win probability + 1 x draw probability) against every other MLS team – both home and away – and taking the average per team.

SEBA has the Union increasing from 16th to 15th.

For those interested in how Philadelphia’s matches are weighted in the model (especially if skeptical about why SEBA’s rankings can be different from other outlets):

‘wght’ is the actual weight value used in the model, which is a combination of the ‘timewght’ (how long ago the match occurred), ‘goalWght’ (how much luck could have influenced the match result, as indicated by the goal differential), and ‘rostWght’ (how similar the roster deployments for both teams were compared with current trends).

By comparison, the current roster expectations for maximum weight for the Union (and therefore the model’s assessment of ‘who’ Philadelphia is in the model) are currently:

The following shows the evolution of SEBA’s power rankings for the MLS Eastern Conference over time:

Playoffs probability and more

Philadelphia’s playoffs odds have remained at 0.3%.

The following shows the simulation distribution for the points earned by the sixth place MLS East club as well as the simulation distribution of points that Philadelphia is expected to earn.

Tiebreakers aside, the Union make the playoffs when >= this MLS Eastern Conference 6th place value.

The most common number of points required to make the playoffs in the East’s 6th slot increased from 48 to 49 while the most common number of points simulated for the Union has remained at 40.

In part, clubs that score a lot of goals are given an advantage in MLS Cup due to the two-leg aggregate goal format of the conference semifinals and finals. That gives those clubs a better chance at notching large victories which carry over.

The Union’s chances of winning the MLS Cup remain at practically zero.

Philadelphia’s chances for qualifying in 2017 for the 2019 edition of the CONCACAF Champions League remain at practically zero.

Over time, we can see how Philadelphia’s odds for different prizes have changed:

The following shows the probability of each playoff ranking finish:

The following shows the summary of the simulations in an easy table format:

Next, we show how the Remaining Strength of Schedule affects each team.

The “Points Percentage Advantage” shown on the X-axis represents the percentage of points expected over the league average schedule. This “points expected” value is generated by simulating how all teams would perform with all remaining schedules (and therefore judges a schedule based upon how all teams would perform in that scenario).

In short, the higher the value, the easier the remaining schedule.

Accompanying the advantage percentage in the following table is their current standings rank (right now ties are not properly calculated beyond pts/gd/gf; I may fix that, but maybe not for a while), the remaining home matches, the remaining away matches, the current average points-per-game of future opponents (results-based, not model-based), and the average power ranking of future opponents according to SEBA.

The following shows the expectations for upcoming Philadelphia matches:

Last Game Probability Chart

This model finally incorporates changing team statistics due to subs and yellow/red cards.

For the following, the green line represents the odds of a win, the blue line the odds of a tie, and the red line the odds of a loss.

The following shows the changing proportion of Philadelphia’s probability of scoring goals compared with their opponents’. This proportion can only change due to subs, yellow cards, and red cards.

*For example, a value of 2 means that Philadelphia is twice as likely to score as its opponent. A value of 0.5 means that Philadelphia is half as likely to score as its opponent.

The following shows the changing raw probability of the two teams each scoring a goal. Green is Philadelphia’s probability of scoring a goal and red is their opponent’s probability of scoring a goal.

The reason for the spike at the 45th minute is because ’45+x’ is condensed to the 45th minute (therefore increasing the frequency of goals occurring in the 45th minute) to avoid duplication with the actual 46th/47th/etc minute, whereas the same situation does not occur for ’90+x’ minute, for which we actually calculate the addition and attribute the action to the goals to the 91st/92nd/etc minute if they occur.

 

 Philadelphia +/- Player Analysis

The ‘+’ is a measure counting how many goals were scored by the Union while the player was on the field. The ‘-‘ counts how many goals were scored against the Union while the player was on the field.

The first table is for 2017. The second table is for all-time since 2013.

Player Net + MINS Net/90 +/90 -/90
1 Chris Pontius 10 32 22 4359 0.206 0.661 0.454
2 Jack Elliott 6 35 29 5654 0.096 0.557 0.462
3 Oguchi Onyewu 5 29 24 4770 0.094 0.547 0.453
4 Warren Creavalle 4 8 4 874 0.412 0.824 0.412
5 Giliano Wijnaldum 3 19 16 3087 0.087 0.554 0.466
6 Andre Blake 3 30 27 5130 0.053 0.526 0.474
7 CJ Sapong 2 39 37 6624 0.027 0.530 0.503
8 Fabian Herbers 1 8 7 635 0.142 1.134 0.992
9 Raymon Gaddis 1 26 25 4530 0.020 0.517 0.497
10 Alejandro Bedoya 1 33 32 5817 0.015 0.511 0.495
11 Haris Medunjanin 0 40 40 7191 0.000 0.501 0.501
12 Ilsinho 0 22 22 3160 0.000 0.627 0.627
13 Derrick Jones 0 10 10 1534 0.000 0.587 0.587
14 Fabinho -2 21 23 3956 -0.046 0.478 0.523
15 Fafa Picault -2 24 26 3833 -0.047 0.564 0.610
16 Keegan Rosenberry -2 14 16 2437 -0.074 0.517 0.591
17 Roland Alberg -2 14 16 1693 -0.106 0.744 0.851
18 Adam Najem -2 1 3 142 -1.268 0.634 1.901
19 John McCarthy -3 10 13 2070 -0.130 0.435 0.565
20 Joshua Yaro -4 5 9 942 -0.382 0.478 0.860
21 Marcus Epps -5 5 10 824 -0.546 0.546 1.092
22 Jay Simpson -5 4 9 350 -1.286 1.029 2.314
23 Richie Marquez -6 11 17 2313 -0.233 0.428 0.661
Player Net + MINS Net/90 +/90 -/90
1 Conor Casey 9 64 55 8296 0.098 0.694 0.597
2 Chris Pontius 8 76 68 11706 0.062 0.584 0.523
3 Jack Elliott 6 35 29 5654 0.096 0.557 0.462
4 Antoine Hoppenot 5 21 16 1060 0.425 1.783 1.358
5 Oguchi Onyewu 5 29 24 4770 0.094 0.547 0.453
6 Jack McInerney 5 35 30 5403 0.083 0.583 0.500
7 Fred 3 8 5 689 0.392 1.045 0.653
8 Giliano Wijnaldum 3 19 16 3087 0.087 0.554 0.466
9 Brian Brown 2 6 4 328 0.549 1.646 1.098
10 Warren Creavalle 2 40 38 6025 0.030 0.598 0.568
11 Ethan White 2 36 34 6056 0.030 0.535 0.505
12 Matt Kassel 1 2 1 139 0.647 1.295 0.647
13 Matthew Jones 1 3 2 450 0.200 0.600 0.400
14 Gabriel Farfan 1 6 5 643 0.140 0.840 0.700
15 Fabian Herbers 1 34 33 3686 0.024 0.830 0.806
16 Ilsinho 1 48 47 6333 0.014 0.682 0.668
17 Haris Medunjanin 0 40 40 7191 0.000 0.501 0.501
18 Michael Lahoud 0 36 36 5294 0.000 0.612 0.612
19 Keon Daniel 0 22 22 3481 0.000 0.569 0.569
20 Carlos Valdés 0 11 11 1946 0.000 0.509 0.509
21 Derrick Jones 0 10 10 1534 0.000 0.587 0.587
22 Bakary Soumaré 0 3 3 504 0.000 0.536 0.536
23 Zac MacMath -1 88 89 15930 -0.006 0.497 0.503
24 Vincent Nogueira -1 81 82 13892 -0.006 0.525 0.531
25 Brian Sylvestre -1 18 19 3330 -0.027 0.486 0.514
26 Austin Berry -1 9 10 1667 -0.054 0.486 0.540
27 Rais Mbolhi -1 1 2 270 -0.333 0.333 0.667
28 Walter Restrepo -1 3 4 162 -0.556 1.667 2.222
29 Roger Torres -1 1 2 68 -1.324 1.324 2.647
30 Corben Bone -1 0 1 12 -7.500 0.000 7.500
31 Jeff Parke -2 37 39 6795 -0.026 0.490 0.517
32 Fafa Picault -2 24 26 3833 -0.047 0.564 0.610
33 Rais M’bolhi -2 9 11 1800 -0.100 0.450 0.550
34 Pedro Ribeiro -2 4 6 611 -0.295 0.589 0.884
35 Anderson Conceicão -2 0 2 180 -1.000 0.000 1.000
36 Adam Najem -2 1 3 142 -1.268 0.634 1.901
37 Charlie Davies -2 2 4 52 -3.462 3.462 6.923
38 Raymond Lee -2 0 2 24 -7.500 0.000 7.500
39 Sheanon Williams -3 95 98 17177 -0.016 0.498 0.513
40 Alejandro Bedoya -3 46 49 8349 -0.032 0.496 0.528
41 Leonardo Fernandes -3 4 7 690 -0.391 0.522 0.913
42 Cristián Maidana -4 62 66 10287 -0.035 0.542 0.577
43 Tranquillo Barnetta -4 54 58 9477 -0.038 0.513 0.551
44 Ken Tribbett -4 33 37 5681 -0.063 0.523 0.586
45 Michael Farfan -4 29 33 4822 -0.075 0.541 0.616
46 Joshua Yaro -4 25 29 4288 -0.084 0.525 0.609
47 CJ Sapong -5 108 113 18518 -0.024 0.525 0.549
48 Amobi Okugo -5 87 92 16050 -0.028 0.488 0.516
49 Kléberson -5 6 11 1303 -0.345 0.414 0.760
50 Marcus Epps -5 5 10 824 -0.546 0.546 1.092
51 Jay Simpson -5 4 9 350 -1.286 1.029 2.314
52 Andre Blake -6 85 91 15840 -0.034 0.483 0.517
53 Keegan Rosenberry -6 67 73 12319 -0.044 0.489 0.533
54 Eric Ayuk -6 19 25 3165 -0.171 0.540 0.711
55 Aaron Wheeler -6 9 15 1730 -0.312 0.468 0.780
56 Sébastien Le Toux -7 119 126 19494 -0.032 0.549 0.582
57 Roland Alberg -8 36 44 4493 -0.160 0.721 0.881
58 Maurice Edu -9 71 80 13587 -0.060 0.470 0.530
59 Danny Cruz -9 47 56 7151 -0.113 0.592 0.705
60 Zach Pfeffer -9 15 24 2338 -0.346 0.577 0.924
61 Fabinho -10 136 146 24514 -0.037 0.499 0.536
62 John McCarthy -10 24 34 5220 -0.172 0.414 0.586
63 Leo Fernandes -10 12 22 2082 -0.432 0.519 0.951
64 Fernando Aristeguieta -12 13 25 2909 -0.371 0.402 0.773
65 Richie Marquez -13 90 103 17096 -0.068 0.474 0.542
66 Steven Vitória -13 19 32 4410 -0.265 0.388 0.653
67 Andrew Wenger -14 53 67 8723 -0.144 0.547 0.691
68 Raymon Gaddis -20 151 171 28749 -0.063 0.473 0.535
69 Brian Carroll -21 119 140 22471 -0.084 0.477 0.561

 

Model Validation

The following shows the overall net values since 2013 which is when data is available.

The following shows the degree of error by the model vs the error if the model was purely random without intelligence. The x-axis is based on the date from which the forecast was made (this will update throughout the season as more results are finalized and compared with predictions). The ordinal squared error metric (not a traditional metric) is calculated as:

(ProbW – ActW)^2 + (ProbT – ActT)^2 + (ProbL – ActL)^2 +

((ProbW + ProbT) – (ActW + ActT))^2 +

((ProbL + ProbT) – (ActL + ActT))^2

where Prob[W/T/L] is the model’s probability of resulting outcomes and Act[W/T/L] is a 1 or 0 representation of whether it actually happened.

Random errors will decline when more ties occur as there is a less severe penalty for ties.

We should expect random errors to remain relatively constant over time, where our model’s errors will hopefully decline as the season goes on as it gathers new information.

These data points are not fixed until the end of the season due to additional matches adding to them.

USL
Power Rankings

SEBA has the Bethlehem Steel climbing from 17th to 15th while it has Harrisburg City remaining at 25th.

The following shows the evolution of SEBA’s power rankings for the USL East over time.

Playoffs probability and more

Bethlehem’s playoff odds have increased from 71.1% to 85.5% while Harrisburg City’s odds of reaching the postseason decreased from 0.1% to practically zero.

Bethlehem’s odds at becoming the USL Champion increased from 0.9% to 1.1% while Harrisburg City’s chances remains at practically zero:

Over time, we can see how the odds for different prizes change for Bethlehem and Harrisburg.

The following shows the probability of each post-playoff ranking finish:

The following shows the summary of simulations in an easy table format.

We can also show how the Remaining Strength of Schedule affects each team.

The “Points Percentage Advantage” shown on the X-axis represents the percentage of points expected over the league average schedule. This “points expected” value is generated by simulating how all teams would perform with all remaining schedules (and therefore judges a schedule based upon how all teams would perform in that scenario).

In short, the higher the value, the easier the remaining schedule.

Remaining home field advantage will make a large contribution here. It can also be true that a better team has an ‘easier’ schedule simply because they do not have to play themselves. Likewise, a bad team may have a ‘harder’ schedule because they also do not play themselves.

The table following the chart also shares helpful context with these percentages.

Accompanying the advantage percentage in the following table is their current standings rank (right now ties are not properly calculated beyond pts/gd/gf), the remaining home matches, the remaining away matches, the current average points-per-game of future opponents (results-based, not model-based), and the average power ranking of future opponents according to SEBA.

The following shows the expectations for upcoming matches for both Bethlehem and Harrisburg:

Model Validation

This chart is the same as that in the MLS forecast (except for USL matches instead of MLS).

Remember that these data points are not fixed until the end of the season.

NWSL
Power Rankings

 

Playoffs probability and more

The following shows the summary of the simulations in an easy table format.

As a new feature, we can also show how the Remaining Strength of Schedule affects each team.

The “Points Percentage Advantage” shown on the X-axis represents the percentage of points expected over the league average schedule. This “points expected” value is generated by simulating how all teams would perform with all remaining schedules (and therefore judges a schedule based upon how all teams would perform in that scenario).

In short, the higher the value, the easier the remaining schedule.

It can also be true that a better team has an ‘easier’ schedule simply because they do not have to play themselves. Likewise, a bad team may have a ‘harder’ schedule because they also do not play themselves.

The table following the chart also shares helpful context with these percentages:

Accompanying the advantage percentage in the following table is their current standings rank (right now ties are not properly calculated beyond pts/gd/gf), the remaining home matches, the remaining away matches, the current average points-per-game of future opponents (results-based, not model-based), and the average power ranking of future opponents according to SEBA.

The SEBA Projection System is an acronym for a tortured collection of words in the Statistical Extrapolation Bayesian Analyzer Projection System. Check out the first season’s post to find out how it works (https://phillysoccerpage.net/2017/03/03/2017-initial-seba-projections/)

2 Comments

  1. As every week, thanks, Chris.

Leave a Reply

Your email address will not be published. Required fields are marked *

*