International Simulation Football League
*Analyzing Win Probability in DDSPF21 - Printable Version

+- International Simulation Football League (https://forums.sim-football.com)
+-- Forum: Community (https://forums.sim-football.com/forumdisplay.php?fid=5)
+--- Forum: Media (https://forums.sim-football.com/forumdisplay.php?fid=37)
+---- Forum: Graded Articles (https://forums.sim-football.com/forumdisplay.php?fid=38)
+---- Thread: *Analyzing Win Probability in DDSPF21 (/showthread.php?tid=44048)

Pages: 1 2

*Analyzing Win Probability in DDSPF21 - infinitempg - 03-06-2023

Analyzing Win Probability in DDSPF21
An exploration of statistics and choking

This has been a project that has been in the works for quite a long time - I actually originally started working on this a bit after the sim transfer with data from the old sim - but I couldn't quite figure out all the weird bugs. A long time later, I came back to the project with the intent of looking at win probabilities in our wonderful current sim engine of DDSPF21. I haven't quite gotten all the weirdness out yet (for example, I don't count kickoffs as plays even though you can score on kickoff returns - because I'm lazy), but I feel like it's in an acceptable enough state that I can make an article out of it.

What is a win probability chart? Well, I'm sure you've seen many of these in the form of the classic Atlanta Falcons choke:

[Image: EiYsqtJWsAAP261.png]

These charts are designed to calculate a team's probability of winning the game based on a few factors, including the current score, who has possession, what the down and distance are, the distance to the goal line, and how much time is left in the game. In the example above we see that with about 2:52 left in the game, Dallas pinned at their own 24 yard line, and a 9 point lead, ESPN claims that the Falcons have a 99.9% probability of winning the game. We can see that prior to this, the Falcons appear to have a solid control of the game with their Win Probability hovering at or above 75%, before it all comes tumbling down.

This of course leads me to show an example of an ISFL Win Probability Chart:

[Image: image.png]

(full size here)

This is a win probability chart for the S40 W6 match between OCO and SJS. As we can see, it was fairly back-and-forth for a majority of the game. Interestingly towards the end we see that SJS actually drags OCO's win probabilty down from 95% all the way to 45% (you can't really see it on the chart though) in about 90 seconds. What's up with that? Let's look at the play by play:

Code:
Q4 2:00 - 2-2 @ SJS 2
OCO 22-25 SJS (SJS WP - 22.99%)
Rush by Zane Cold for 2 yds. TOUCHDOWN!  (SJS WP - 8.21%)

Q4 1:54 - 1-10 @ SJS 10
OCO 29-25 SJS (SJS WP - 11.66%)
Farrell, O. is SACKED by Boone McCoy - DE for -6 yds. (SJS WP - 5.32%)

Q4 1:42 - 2-16 @ SJS 4
OCO 29-25 SJS (SJS WP - 5.32%)
Pass by Farrell, O., complete to Penne, D. for 44 yds. Tackle by Washington, K.. First Down! Timeout called by SJS. (SJS WP - 30.99%)

Q4 1:20 - 1-10 @ SJS 48
OCO 29-25 SJS (SJS WP - 30.99%)
Pass by Farrell, O. to King III, C. falls incomplete. (SJS WP - 25.37%)

Q4 1:15 - 2-10 @ SJS 48
OCO 29-25 SJS (SJS WP - 25.37%)
San Jose Penalty on Peterson, G.: False Start. (SJS WP - 19.20%)

Q4 1:15 - 2-15 @ SJS 43
OCO 29-25 SJS (SJS WP - 19.20%)
Pass by Farrell, O., complete to Penne, D. for 11 yds. Tackle by Schwarzer, R.. Timeout called by SJS. (SJS WP - 26.53%)

Q4 1:08 - 3-4 @ OCO 45
OCO 29-25 SJS (SJS WP - 26.53%)
Play nullified by San Jose Penalty on Bot, B.: Illegal use of Hands. (SJS WP - 15.01%)

Q4 1:01 - 3-14 @ SJS 44
OCO 29-25 SJS (SJS WP - 15.01%)
Pass by Farrell, O., complete to King III, C. for 7 yds. Tackle by LeBayers, J.. Timeout called by SJS. (SJS WP - 12.70%)

Q4 0:55 - 4-6 @ OCO 48
OCO 29-25 SJS (SJS WP - 12.70%)
Pass by Farrell, O., complete to Higgins, J. for 7 yds. Tackle by Rodriguez, B.. First Down! (SJS WP - 38.35%)

Q4 0:43 - 1-10 @ OCO 41
OCO 29-25 SJS (SJS WP - 38.35%)
Pass by Farrell, O., complete to Higgins, J. for 11 yds. Tackle by Jazz, S.. First Down! (SJS WP - 46.98%)

Q4 0:43 - 1-10 @ OCO 29
OCO 29-25 SJS (SJS WP - 46.98%)
Owen Farrell spikes the ball to stop the clock. (SJS WP - 40.52%)

Q4 0:27 - 2-10 @ OCO 29
OCO 29-25 SJS (SJS WP - 40.52%)
Pass by Farrell, O., complete to Higgins, J. for 11 yds. Tackle by LeBayers, J.. First Down! (SJS WP - 55.71%)

Q4 0:27 - 1-10 @ OCO 17
OCO 29-25 SJS (SJS WP - 55.71%)
Owen Farrell spikes the ball to stop the clock. (SJS WP - 49.28%)

Q4 0:20 - 2-10 @ OCO 17
OCO 29-25 SJS (SJS WP - 49.28%)
Pass by Farrell, O., complete to Camacho, P. for 6 yds. Tackle by McAlister, A.. (SJS WP - 48.76%)

Q4 0:20 - 3-3 @ OCO 11
OCO 29-25 SJS (SJS WP - 48.76%)
Owen Farrell spikes the ball to stop the clock. (SJS WP - 26.61%)

Q4 0:11 - 4-3 @ OCO 11
OCO 29-25 SJS (SJS WP - 26.61%)
Pass by Farrell, O., complete to Camacho, P. for 8 yds. Tackle by McAlister, A.. First Down! (SJS WP - 0.00%)

Following this, we see that SJS actually comes incredibly close to scoring the game winning touchdown - coming just 3 yards short! The model actually gives SJS a 55.71% chance of winning with about 27 seconds left - though it is penalized for spiking the ball and giving up 1st down. Obviously, this is the correct thing to do (the clock is running!), but the model only knows that they've wasted a down and now only have 3 opportunities to get at least 10 yards instead of 4.

How does it know all this? We need to talk about something that's been covered recently in the league to learn this...

Expected Points

Expected Points, or EP, are a measure of the amount of points a team is "expected" to score based on the down, distance, and yards to endzone. This topic is covered quite well by @aeonsjenni in her article here, and I highly recommend you check out her work with EP and EPA (expected points added)!

That being said, my method of obtaining an EP model is a bit different than hers. While Jenni did a calculation of average drive result from each down, distance, and yards to goal - I used a generalized linear regression of this data - specifically from only S27, S31, and S35. This resulted in my graphs being a little bit smoother than hers.

[Image: image.png]

[Image: chart_1.png]

The larger trends remain the same, although my data is significantly less dramatic in terms of swing and negative points. I can only assume the smoothness arises from the fact that I'm taking a linear regression and so am more protected from outliers. I don't have a good explanation for why my model doesn't show a strong propensity for negative points at the moment. Perhaps it is based on the way I treat turnovers?

Regardless, we can now use these EP numbers to contextualize entire games, and create a Win Probability model!

WP Model - The Statistical Analysis

In order to make a win probability model, we need to use the following formula:
Code:
WP = 1 - CDF(0,μ,σ)
where the CDF is the cumulative distribution function of a normal (Gaussian) distribution, 0 represents the point spread at the beginning of the game (which better be zero), μ is the average point spread for home teams, and σ is the average standard deviation in points scored.

Between S27 and S39, the home team on average wins by μ = 2.8494 points, and the standard deviation in total points scored is σ = 15.3726 points. Based on this, the home team's win probability at the beginning of the game is about 57.4%. Note, I'm not accounting for relative team strength like ESPN and other WP models do - because I'm lazy.

But what about during a game? We can alter the formula to account for the time in seconds remaining in the game, t:
Code:
WP = 1 - CDF([away points - home points], μt/60, σ√(t/60))

Let's take an example. Suppose the away team is ahead by 10 points at the start of the second quarter. Plugging into this formula, we get that the home team's win probability is about 27.7%! If the home team was instead ahead by 10 at this point in the game, they'd have a 66.4% win probability.

Finally, we need to account for the possessing team's field position - and here is where we will use Expected Points! We can calculate the home team's EP easily (if the away team has the ball, we take the negative of the away team's EP), and then add this EP to our point spread calculation:
Code:
WP = 1 - CDF([away points - home points - homeEP], μt/60, σ√(t/60))
(We subtract homeEP here because we're subtracting the quantity [homePoints + homeEP].)

As an example, the EP for a team on 1st and 10 at midfield is about 3.148 points. If the home team has the ball in the spot, down 10, at the beginning of the second quarter, their win probability is 36.1% - an increase in 8.4%!

Since I have the play-by-play data scraped and smooshed into a Pandas dataframe, I can feed in every play's down, distance, yards to goal, time remaining, and score difference to obtain win probabilities for every play in every game since S27!

The Greatest Chokes

No one pays attention to win probability charts until a team manages to choke away a lead that was thought insurmountable. This is why the most commonly shared charts have to do with the Falcons - it's funnier that way (unless you're a Falcons fan).

This of course leads me to go on a search for the greatest chokes in ISFL history. Of course, one might think of the great double choke in the S33 Ultimus, where OCO blew both an early 24-0 lead and a 38-24 lead with 5 minutes left in the game. How does this stack up?

[Image: image.png]

As expected, Orange County has definitive control for a large majority of the game. Orange County actually raced out to a 95% WP when they took a 24-0 lead. Baltimore does make some headway going into halftime with a 16.3% WP (83.7% OCO WP).

Orange County actually manages to get up to a 99.5% WP with 13:42 to go in the 4th Quarter, as they have the ball on the BAL 5 up by 14. A touchdown there would have certainly iced the game, but instead Carter throws an interception that allows Baltimore to get back within 7. The infamous 38-24 scoreline referenced above is at a point where OCO has a 99.2% WP - though this is bumped down to 95.1% when Baltimore scores.

This of course is one of the most legendary chokes in league history, but surprisingly it is far from the worst. A search of games where a team had a WP of greater than 99.9% before losing returns a list of 15 games! I will give a description of the top 5 games.

In descending order:

15. S39 W12 SAR @ CTC

Highest Loser WP: 99.904%

Game Log: https://index.sim-football.com/ISFLS39/Boxscores/Boxscore.html?id=5194

[Image: image.png]
(full size here)

14. S29 W8 CHI @ HON

Highest Loser WP: 99.943%

Game Log: https://index.sim-football.com/ISFLS29/Boxscores/Boxscore.html?id=1034

[Image: image.png]
(full size here)

13. S27 W16 COL @ HON

Highest "Loser" WP: 99.951%

Game Log: https://index.sim-football.com/ISFLS27/Boxscores/Boxscore.html?id=263

[Image: image.png]
(full size here)

12. S33 W15 CHI @ YKW

Highest Loser WP: 99.963%

Game Log: https://index.sim-football.com/ISFLS33/Boxscores/Boxscore.html?id=2737

[Image: image.png]
(full size here)

11. S29 W9 NYS @ SAR

Highest Loser WP: 99.974%

Game Log: https://index.sim-football.com/ISFLS29/Boxscores/Boxscore.html?id=1039

[Image: image.png]
(full size here)

10. S35 W15 SAR @ OCO

Highest Loser WP: 99.981%

Game Log: https://index.sim-football.com/ISFLS35/Boxscores/Boxscore.html?id=3560

[Image: image.png]
(full size here)

9. S29 W2 SAR @ BER

Highest Loser WP: 99.989%

Game Log: https://index.sim-football.com/ISFLS29/Boxscores/Boxscore.html?id=988

[Image: image.png]
(full size here)

8. S39 W17 NOLA @ AUS

Highest Loser WP: 99.991%

Game Log: https://index.sim-football.com/ISFLS39/Boxscores/Boxscore.html?id=5224

[Image: image.png]
(full size here)

7. S29 W2 YKW @ BAL

Highest Loser WP: 99.996%

Game Log: https://index.sim-football.com/ISFLS29/Boxscores/Boxscore.html?id=989

[Image: image.png]
(full size here)

6. S28 W12 SAR @ COL

Highest Loser WP: 99.997%

Game Log: https://index.sim-football.com/ISFLS28/Boxscores/Boxscore.html?id=645

[Image: image.png]
(full size here)

And now, the Top 5:

5. S40 W3 SAR @ BAL

Highest Loser WP: 99.997%

Game Log: https://index.sim-football.com/ISFLS40/Boxscores/Boxscore.html?id=5538

[Image: image.png]
(full size here)

We start our analysis with an entry from this season! Sarasota stormed off to a 31-10 lead over Baltimore with a pick-six in the mid-3rd quarter. A 21 point lead is a pretty strong lead, and that's represented by a 98.35% WP for Sarasota. However, we're still in the third quarter - and Baltimore manages to get two touchdowns (one being their own pick-six!) back to bring the score within 7 before the quarter is even over (SAR WP: 78.54%). Sarasota then manage to score two field goals over the next 7 minutes to bring their WP up to 98.71% with 8:08 remaining.

The teams trade punts for two possessions, giving Baltimore the ball back with 5:34 left (SAR WP: 99.77%). Baltimore starts to drive and actually makes it to the SAR 36 before being hit by an Illegal use of Hands penalty to force 2nd and 20 with 2:45 left (SAR WP: 99.997%). They dig themselves out of this with a 23 yard pass, before converting a 3rd and 10 a minute later to themselves into a goal-to-go scenario. Dante King's 8 yard receiving touchdown with 1:33 left allowed the Hawks to close the lead down to 6 (SAR WP: 98.23%). Unfortunately for Sarasota, Baltimore manage to collect the onside kick as well (SAR WP: 89.35%).

After a quick first down via penalty, Preston Beatz takes a crushing 11 yard sack, a +12.81% WP swing for the Sailfish taking their WP up to 92.83%. They do get 5 yards back, but the Hawks find themselves in a do-or-die 4th and 16 at the SAR 38, with 42 seconds remaining (SAR WP: 96.10%) - but they manage to hit another huge passing play for the touchdown! As they get the ball back on their own 25 with 0:25 left, Sarastoa's WP has dropped down to 40.92% - to be honest, surprisingly high given they have to drive at least 40 yards to get into field goal range. They do manage to get 16 of those yards, but it's too little, too late.

4. S30 W2 SAR @ BER

Highest Loser WP: 99.998%

Game Log: https://index.sim-football.com/ISFLS30/Boxscores/Boxscore.html?id=1401

[Image: image.png]
(full size here)

While it was a fairly close game for the most part, Saleem Spence's touchdown with 9:18 to go to put the Sailfish up 13 looked like it would set a pretty solid framework for them to get the win (SAR WP: 98.26%). Things continued to improve for the Sailfish as Berlin, despite driving effectively down the field, was killing tons of time doing so. Berlin's lowest point came with 4:03 left in the game, when on 3rd and 3 they threw an incomplete pass - putting Sarasota's WP up to it's highest point of 99.998%. With 4th and 3 at their opponent's 33, Berlin was in do-or-die mode as a field goal would not help them improve their odds. They go for the 4th down and manage to convert it, as Troy Abed gains 13 yards on the catch and run to put them at the SAR 20 with 3:29 to go (SAR WP: 99.91%). The Fire Salamanders convert another first down two plays later, though they take two sacks in three plays in a goal-to-go situation to be stuck with 4th and Goal, 21 yards out, with 2 minutes to go (SAR WP: 99.997%). Despite this, Kaepercolin manages to find a touchdown pass and bring the gap down to 6 (SAR WP: 98.23%).

Sarasota get the ball back and immediately run the ball three straight times, forcing the Fire Salamanders to burn all three of their timeouts. However, they only manage to get to 4th and inches at their own 34 (SAR WP: 98.36%), and in a perhaps questionable decision decide to punt the ball away. This punt only goes 38 yards, meaning that Berlin now have the ball at their own 27 with 1:31 remaining (SAR WP: 94.45%).

Immediately, Berlin complete a 38 yard pass to bring them to the Sarasota 34 (SAR WP: 81.02%) before another quick 9 yard pass (SAR WP: 74.25%) takes them down to the minute mark. On the next play, Kaepercolin finds Tychondrius Hood in the endzone for a 24 yard touchdown, and the extra point gives the Fire Salamanders an unlikely lead! At this point, Sarasota's WP sits at 34.82%, meaning they lost nearly 60% WP in the span of 1 minute. That being said, the Sailfish still have a chance with 39 seconds left - though they only manage to get to their own 47 before running out of time.

3. S29 W13 PHI @ NYS

Highest Loser WP: 99.998%

Game Log: https://index.sim-football.com/ISFLS29/Boxscores/Boxscore.html?id=1068

[Image: image.png]
(full size here)

This is a close game for most of the first three quarters, but New York starts to pull away as the fourth quarter begins - taking an 11 point lead with a 49 yard touchdown pass. Philly does manage to get 3 points back with 7:56 left in the game (NYS WP: 93.39%) - but New York kills another 4 minutes of clock and kicks their own field goal to extend the lead back to 11 with 3:04 left (NYS WP: 99.996%). This is when things start to get interesting.

Philadelphia starts off at their own 11 after a short kickoff return (NYS WP: 99.990%), and manage to convert a 3rd and 3 in order to get out of the hole with 2:30 to go (NYS WP: 99.993%). And then they immediately go back in the hole as Negs takes an 11 yard sack to put them back at their own 11. This is where NYS has their highest WP: 99.998%.

And then it all turns upside down. Ryan Negs throws an 88 yard touchdown to Flash Panda, and then they convert the two point conversion to bring the game within 3 (NYS WP: 86.49%). Then, Philly successfully recovers the onside kick (NYS WP: 58.29%) to take over at their own 48 yard line. Immediately after, Ryan Negs throws another 52 yard touchdown to take a 4 point lead! In the span of two plays, an onside kickoff, and a 2 point conversion, New York has gone from a 99.998% WP all the way down to an 8.21% WP. NYS get the ball back with about 30 seconds to go, but they don't have the same magic that Negs had and fail to even make it past midfield.

2. S36 W16 NYS @ AZ

Highest Loser WP: 99.999%

Game Log: https://index.sim-football.com/ISFLS36/Boxscores/Boxscore.html?id=3980

[Image: image.png]

(full size here)

Arizona holds a 16 point lead with just under 5 minutes to go - not an iron-clad lead but certainly one that most would be comfortable with. However, New York has spent the last 7 minutes of game time slowly and methodically making their way down the field. With 4:34 left in the game (AZ WP: 99.54%), New York finally punch the ball into the endzone - though they fail the 2 point conversion. Arizona's WP drops down to 99.28%, and of course they get the ball as well.

Arizona then proceeds to complete a quick pass before Jay Cue Jr. runs for a first down - hitting their maximum WP of 99.99889% with 3:04 to go. Over the next three plays, New York manages to force a 4th and 1 at the Arizona 42 (AZ WP: 99.9908%). The Outlaws choose to try and end the game right then and there, but Cue cannot convert and they turn the ball over on downs (AZ WP: 99.433%).

With Savage immediately taking a 6 yard sack, it looks like Arizona may still escape. Even better for Arizona, they manage to force a tough 4th and 6 for New York at the AZ 40 with 1:38 left (AZ WP: 99.94%). Given the choice between a 57 yard field goal and converting 6 yards, the Silverbacks risk it for the biscuit and manage to get 9 yards! This drops Arizona's WP down to 98.92% - still good, but definitely nerve-wracking now.

New York manages to convert a tough 3rd and 4 before ending up with a 4th and 3 at the AZ 13 with 0:50 left (AZ WP: 99.78%). Faced with the same decision, New York goes for it again and score the touchdown! Now, Arizona's WP falls down to 86.49%.

BUT - New York manage to convert the onside kick! Suddenly, Arizona's WP drops down to 57.60% and Outlaw fans are definitely terrified now. The Silverbacks drive into field goal range and with 16 seconds left they convert the 55 yard field goal to tie the game! Overtime goes back and forth for a bit, but New York manages to emerge victorious in the end.

1. S33 W6 CHI @ COL

Highest Loser WP: 99.999997%

Game Log: https://index.sim-football.com/ISFLS33/Boxscores/Boxscore.html?id=2670

[Image: image.png]
(full size here)

This is the undisputed greatest choke of DDSPF21 history. I'm sure there might be one of this quality in the old sim, but I'm too lazy to find out. Colorado had a 21 point lead up until 4:11 remaining in the fourth quarter. Prior to Chicago's touchdown, the Yeti had a WP of 99.9957%, and Zohri's touchdown only decreased the WP to 99.9810% - because if we're being honest, blowing a 14 point lead in less than 5 minutes (nevermind the full 21) is incredibly unrealistic.
And yet. Colorado immediately takes the ball and plays extremely aggressive. On 1st and 10 (COL WP: 99.9962%), they throw an incompletion. On 2nd an 10 (COL WP: 99.9999971%, the highest WP they have, interestingly!), they throw an incompletion. On 3rd and 10 (COL WP: 99.999919%), they throw an incompletion. The Yeti burn off just 39 seconds on this possession, never forcing Chicago to burn any of their timeouts.

Chicago then proceed to march down the field over 2 minutes (using 1 of those timeouts and converting a 4th and 6) before scoring a touchdown with 1:38 left in the game, knocking the Yeti lead down to 7 and the Yeti WP down to 99.4263%. On the next possession, Colorado plays a little bit more conservative with a quick pass and then two rushes - though they fail to get the first down. Thanks to their 2 remaining timeouts, Chicago reduce the time loss to 30 seconds.

With 4th and 1 at their own 33 (COL WP: 99.4494%), the Yeti opt to punt. This is fundamentally a good decision, but the punt only goes 33 yards - and then is returned another 15 yards! This is a net gain of 18 yards in field position, and reduces the COL WP to 94.89%. Chicago take advantage of the short field position and manage to complete the impossible comeback, tying the game with 0:11 seconds left. Much like Super Bowl LI, Chicago get the ball first in overtime and never let Colorado amend for their mistakes.


I hope this was an illuminating look into the statistics and analysis that goes into generating a win probability chart! It's always fun to laugh at chokes, but it is kind of wild to see some of the completely blown leads that the sim allows for. There has always been the conspiracy that the sim comes up with the result it wants before actually simulating the game result, and sometimes you have to look at these play-by-plays and think there really is some credibility to it. I don't think any NFL or college football team is going to come back from down 21 points with 4:11 left on the clock - but maybe I'm wrong and it has happened. It's just so unrealistic.

I have hinted a little bit about the decision making for fourth downs, and if you've been on NFL Twitter I'm sure you've seen the great fourth down bot by Ben Baldwin. I actually have this made already and have been referencing it while writing the choke analysis pieces - but I don't anticipate I will have the time to write up that report for at least a week or two, if not more. There are some really interesting statistics in there, and definitely some helpful nudges to try and make things more realistic than even the sim goes for.

RE: Analyzing Win Probability in DDSPF21 - Yeenoghu - 03-06-2023

This is super cool. If I might make a request, could you show this game? Surprised it didn't make the cut, but you mentioned kickoff returns acting a little funky...

https://index.sim-football.com/ISFLS28/Boxscores/Boxscore.html?id=583

RE: Analyzing Win Probability in DDSPF21 - Frostbite - 03-06-2023

Why does my ISFL team have to act like the falcons too this is supposed to be my escape

Sad!

RE: Analyzing Win Probability in DDSPF21 - Akoustique - 03-06-2023

Quote:No one pays attention to win probability charts until a team manages to choke away a lead that was thought insurmountable. This is why the most commonly shared charts have to do with the Falcons - it's funnier that way (unless you're a Falcons fan).
[Image: rude-jack-black.gif]

RE: Analyzing Win Probability in DDSPF21 - DREAMSLOTH - 03-06-2023

I am delightfully surprised (and slightly amazed) that Philly / CTC is somehow not on the receiving end of worst chokes of all times.

Question for you about methodology: am I misremembering that the engine affords advantage to home field teams? If those teams do get an advantage, how does DDSPF21's weighting factor into the win probability formula?

RE: Analyzing Win Probability in DDSPF21 - infinitempg - 03-06-2023

(03-06-2023, 07:50 PM)Yeenoghu Wrote: This is super cool. If I might make a request, could you show this game? Surprised it didn't make the cut, but you mentioned kickoff returns acting a little funky...

https://index.sim-football.com/ISFLS28/Boxscores/Boxscore.html?id=583

Interestingly, Pat asked me the same exact question a few days ago.

[Image: image.png]

Notice it got the final score wrong - this is because I just removed all kickoffs lol

Chicago's WP after their FG should be at 96.5%.

(03-06-2023, 08:15 PM)dreamSloth Wrote: I am delightfully surprised (and slightly amazed) that Philly / CTC is somehow not on the receiving end of worst chokes of all times.

Question for you about methodology: am I misremembering that the engine affords advantage to home field teams? If those teams do get an advantage, how does DDSPF21's weighting factor into the win probability formula?

I basically use that weighting as the starting win probability - so I took the straight average of point spread for the home team over the last 12 seasons as well as the standard deviation, and plugged that into the WP formula for t = 0s.

RE: Analyzing Win Probability in DDSPF21 - slate - 03-06-2023

Team records in the 15 games listed at the end of the media:

ARI 0-1
AUS 0-1
BAL 1-1
BER 1-1
CHI 3-0
COL 0-2-1
CTC 2-0 (PHI)
HON 0-1-1
NOLA 1-0
NYS 2-1
OCO 1-0
SJS 0-0
SAR 2-5
YKW 2-0

SAR and COL the only teams to have lost more than one game with 99.9%+ win probability. Sarasota has lost 5.

Sarasota sneakily one of the most sim cursed franchises.

RE: Analyzing Win Probability in DDSPF21 - infinitempg - 03-06-2023

(03-06-2023, 09:27 PM)slate Wrote: Team records in the 15 games listed at the end of the media:

ARI 0-1
AUS 0-1
BAL 1-1
BER 1-1
CHI 3-0
COL 0-2-1
CTC 2-0 (PHI)
HON 0-1-1
NOLA 1-0
NYS 2-1
OCO 1-0
SJS 0-0
SAR 2-5
YKW 2-0

SAR and COL the only teams to have lost more than one game with 99.9%+ win probability. Sarasota has lost 5.

Sarasota sneakily one of the most sim cursed franchises.

they're in almost half the top 15

[Image: 7dj6sq.png]

RE: Analyzing Win Probability in DDSPF21 - aeonsjenni - 03-07-2023

(03-06-2023, 07:38 PM)infinitempg Wrote: Expected Points

Expected Points, or EP, are a measure of the amount of points a team is "expected" to score based on the down, distance, and yards to endzone. This topic is covered quite well by @aeonsjenni in her article here, and I highly recommend you check out her work with EP and EPA (expected points added)!

That being said, my method of obtaining an EP model is a bit different than hers. While Jenni did a calculation of average drive result from each down, distance, and yards to goal - I used a generalized linear regression of this data - specifically from only S27, S31, and S35. This resulted in my graphs being a little bit smoother than hers.

[Image: image.png]

[Image: chart_1.png]

The larger trends remain the same, although my data is significantly less dramatic in terms of swing and negative points. I can only assume the smoothness arises from the fact that I'm taking a linear regression and so am more protected from outliers. I don't have a good explanation for why my model doesn't show a strong propensity for negative points at the moment. Perhaps it is based on the way I treat turnovers?

Regardless, we can now use these EP numbers to contextualize entire games, and create a Win Probability model!


Okay my graph looks really awful here but in my defense this is definitely the most primitive version of it. This is only using 5 seasons of data and has a lot of weird quirks in the data. I made a new one that looks a lot better:

[Image: plot_zoom_png.png]

Beyond that I'm really happy to see this article. It also puts into perspective why the play-by-play data you gave me was formatted the way it was, since obviously you were looking for different information than I was. I find that pretty funny. For some reason I did not expect the home team to have a measurable advantage in the sim, but a +7.4% win probability is really quite significant!

RE: Analyzing Win Probability in DDSPF21 - Michiganonymous - 03-07-2023

The biggest surprise on this list is a team other than Sarasota occupying the top spot.

I’m actually shocked that OCO game is only the 10th-worst choke in league history.

I suppose there’s no way for the model to account for context: it was “win and in” for the Sailfish and the OCO loss knocked us out of the S35 playoffs.