Certified Legendary Thread Race for the flag, in squiggly lines

[SimonH]

Maybe this will just be an odd year - but a three win difference between 7th and 8th? That would surprise me.

As at last night it was an awesome 5 wins! (The 7th placed team was going to be the unluckiest 7th placed team not only in current AFL history, but for the next 100 years as well, by finishing 17:5.) So at this rate, it'll only be a couple more days' tweaking until the 8th ranked team wins more games than the 7th. No, wait...

 

[Final Siren]

Heh, yeah, I was thinking of leaving both versions of the ladder predictor up, and calling the previous one the DRAMATIC LADDER PREDICTOR. Because under that, one mediocre game and suddenly you're finishing five rungs lower.

But nah, I will try to make it accurate.

What it's doing now: It's awarding a percentage of a win each game, based on how likely it thinks it is - e.g. it thinks the chances of Geelong beating Carlton on Friday are 75%, so it's giving Geelong 0.75 for that win and Carlton 0.25. At the end of the H&A season, it rounds everyone off to a whole number of games.

Those percentages come from the predicted margin, since the accuracy of the squiggle's tips can be described fairly reliably (r2=0.89) with LIKELIHOOD OF WIN = MARGIN * 0.01 + 0.52. For example, over the last 30 years, when the squiggle has tipped a team by 10 points, it's been right 62% of the time. The percentage is capped at 97%.

This model reflects the fact that while a team might start favourite in all its games, it's really unlikely to win all of them. For example, consider two games in which your team's chances of winning are 70% each time. You should tip your team both times, of course. But it's actually unlikely to win both, because the odds of that are 0.70 x 0.70 = 0.49, or 49%. This is what happens when one outcome depends upon another: its chances shrink really quickly. And so it's slightly more likely that your team will lose at least one game. (The odds of losing both games are 0.3 x 0.3 = 9%; the odds of losing one or the other are (0.7 x 0.3) + (0.7 x 0.3) = 42%).

 

[clogged]

Final Siren

This person has been running a forecaster for years. Not sure their methodology, but it would be interesting to see whether the two end up agreeing (they haven't plugged in the round 11 results yet).

http://footyforecaster.com/AFL/SeasonForecast

 

[The Swans Blog]

Now we need two clubs to drop down badly enough to make nipples

Haha messed up ****

 

[Wandering Hawk]

...
This model reflects the fact that while a team might start favourite in all its games, it's really unlikely to win all of them. For example, consider two games in which your team's chances of winning are 70% each time. You should tip your team both times, of course. But it's actually unlikely to win both, because the odds of that are 0.70 x 0.70 = 0.49, or 49%. This is what happens when one outcome depends upon another: its chances shrink really quickly. And so it's slightly more likely that your team will lose at least one game. (The odds of losing both games are 0.3 x 0.3 = 9%; the odds of losing one or the other are (0.7 x 0.3) + (0.7 x 0.3) = 42%).

You assume the probabilities are independent, that there is no path dependence. ;-)

You might lose because of critical injuries and then lose again because they are still out. Nitpick Nitpick.

When you model your score distribution about an expected score, why not use a binomial distribution? Chance bounces both ways and thus creates a stronger weighted central tendency.

Z scores and standard deviation?
You could plot how far away teams score from the prediction each weak and create a standard deviation for the season. This would then give you a nice measure to discuss the probability of an upset.

I say you... I have a job ;-)

 

[Final Siren]

You assume the probabilities are independent, that there is no path dependence. ;-)

You might lose because of critical injuries and then lose again because they are still out. Nitpick Nitpick.

When you model your score distribution about an expected score, why not use a binomial distribution? Chance bounces both ways and thus creates a stronger weighted central tendency.

Z scores and standard deviation?
You could plot how far away teams score from the prediction each weak and create a standard deviation for the season. This would then give you a nice measure to discuss the probability of an upset.

I say you... I have a job ;-)

Yeah, but then I'd be building a full-on predictive model. I do like to fiddle with that kind of thing, but it's incredibly geeky and becomes more about mathematics than football.

What I like about the squiggle is it's simple enough to understand, so you can use it to complement your own football knowledge. It's a visualization of things you already know, really. And you can get a sense of where the squiggle might be wrong and right and why. As opposed to a complex mathematical model that basically asks you to trust that it knows better than you do, because you can't understand its inner workings.

I'm here because I like using maths to understand football better, basically, not to use footy to conduct an exercise in maths.

 

[Wandering Hawk]

I'm here because I like using maths to understand football better, basically, not to use footy to conduct an exercise in maths.

Awwws :-(

I seriously doubt that everyone fully follows the smoothing moving averages you are using.

New Offence = 0.91X + 0.09(Y - Z)*A/85

Where X =old offence
Y = your score
Z = the ground modifier
A = their defence

You justify your selection of parameters and formula by appealing to this one being optimal out of an extensive field of modelling of other values...

And then you say you're not into maths modelling that the masses won't follow.

C'mon. Seriously. How do you manage?

 

[Final Siren]

Awwws :-(

I seriously doubt that everyone fully follows the smoothing moving averages you are using.

New Offence = 0.91X + 0.09(Y - Z)*A/85

Where X =old offence
Y = your score
Z = the ground modifier
A = their defence

See, that looks complicated because you mathed it up. But the concept is pretty simple. We have OFF and DEF values for a team. We figure out what OFF and DEF values would have been required to generate the actual results we saw this week. The team shifts towards those values by combining most of its previous values with a little of the new ones.

You justify your selection of parameters and formula by appealing to this one being optimal out of an extensive field of modelling of other values...

And then you say you're not into maths modelling that the masses won't follow.

C'mon. Seriously. How do you manage?

Well, serious modelling is serious business. Increasing amounts of time and smarts are required to eke out each extra point of accuracy, and in the end, how will you ever know how well it's working? The world's greatest model is not going to predict that in two weeks' time Paul Roos will run over a dog and his heart-rending pre-match address will inspire the Demons to a 120-point demolition of North Melbourne. Or that Alastair Clarkson comes down with Guillain-Barr and in his absence the Hawks almost lose to a side that hasn't conceded fewer than 20 goals since Round 4. There isn't enough data to test a sensitive model with confidence, because footy changes all the time: what won games in Round 3 is not what wins games in Round 13. So it sounds like a mug's game to me.

I like answering questions like: "At quarter time, is it better to be 4.0 or 1.8?" And: "How far do you have to be in front before you can feel confident that you're going to win?" Things that help me get how footy works.

Also, drawing squiggles.

 

[d33my]

* I wish I paid more attention in school. when I see it applied like this maths is fascinating but I'm way out of my depth.

 

[Roo Lives Again]

Final Siren

This person has been running a forecaster for years. Not sure their methodology, but it would be interesting to see whether the two end up agreeing (they haven't plugged in the round 11 results yet).

http://footyforecaster.com/AFL/SeasonForecast

Yeah it's pretty s**t though

 

[The Swans Blog]

Yeah it's pretty s**t though

They updated it from round 10 to 11 - Sydney jumped up a few spots!

 

[inbetween]

What it's doing now: It's awarding a percentage of a win each game, based on how likely it thinks it is - e.g. it thinks the chances of Geelong beating Carlton on Friday are 75%, so it's giving Geelong 0.75 for that win and Carlton 0.25. At the end of the H&A season, it rounds everyone off to a whole number of games.

Awarding teams partial wins would result in a more realistic ladder.
But I would advise against rounding the final figure, and simply having the teams with a non-integer win tally. Though the ladder would have most, if not all, teams with a fractional number of wins, this way you can ensure that the total number of wins available in a season remains consistent with the number of games.

If you were to round, you could potentially increase/decrease the total number of wins (league-wide) to a figure that's actually greater/lower than the actual number of games played. A very basic example between a 3 team league to demonstrate this:

Team A (80%) v Team B (20%)
Team A (80%) v Team C (20%)
Team B (50%) v Team C (50%)

Team A: 0.8 + 0.8 = 1.6 wins = 2 wins (rounded)
Team B: 0.2 + 0.5 = 0.7 wins = 1 win (rounded)
Team C: 0.2 + 0.5 = 0.7 wins = 1 win (rounded)

As you can see there were only 3 games played but when rounding, the final ladder would ultimately award a total of 4 wins from the 3 games.
However, the actual ladder itself doesn't use decimal values and partial wins, but with the decimal values, the predictor will actually rank the teams in the correct order. i.e. a 16.9 win team is ranked above a 16.6 team.

 

[Sydney Bloods]

They updated it from round 10 to 11 - Sydney jumped up a few spots!

we're somehow finishing second and geelong 7th?

lol at Hawks only winning 8 more games.

 

[Alfonz]

Final Siren

This person has been running a forecaster for years. Not sure their methodology, but it would be interesting to see whether the two end up agreeing (they haven't plugged in the round 11 results yet).

http://footyforecaster.com/AFL/SeasonForecast

Doesn't make sense ... their weekly predictions (which I presume should be used to forecast the season) have Hawthorn and Fremantle only losing 1 more game each. Which is at odds with the end of season prediction. Didn't check any other sides.

 

[estibador]

**** I wish I paid more attention in school. when I see it applied like this maths is fascinating but I'm way out of my depth.

I don't understand half the posts in this thread but I like looking at the pretty colours on the graphs.

 

[Chunderboy]

Okay, I know this has been explained 1000 times before and I'm sorry, but to make it really simple for me...

It's pretty much, you work a teams projected score based on a combination of their own scoring, and the opponents defence. Then compare this projection to the actual outcome, and a teams offence moves accordingly on the vertical.

Then pretty much do the same for defence and the teams will move on the horizontal.

Is this pretty much right?

 

[SHAKESPEARE]

This last page....

 

[clogged]

Doesn't make sense ... their weekly predictions (which I presume should be used to forecast the season) have Hawthorn and Fremantle only losing 1 more game each. Which is at odds with the end of season prediction. Didn't check any other sides.

I imagine Final Siren's forecaster works the same.

 

[Wandering Hawk]

Okay, I know this has been explained 1000 times before and I'm sorry, but to make it really simple for me...

It's pretty much, you work a teams projected score based on a combination of their own scoring, and the opponents defence. Then compare this projection to the actual outcome, and a teams offence moves accordingly on the vertical.

Then pretty much do the same for defence and the teams will move on the horizontal.

Is this pretty much right?

Yes mate.

Each game we use 90% of the old values plus ten percent of the values that would have produced this week's score.

Worked example.

You have two teams playing, both with fifty offence and defence ratings. I will call them winner and loser.

The expected result is a draw, kicking 85 each.

Actually winner kicks 170 and loser 85.

So looking at winner's new ratings:
New Offence rating is 91% of the old value (50), plus 9% of the offence that would have been needed to kick the 170pts, which is 100 in this case.
This adds up to a new offence of 54.5. Winner moves 4.5 vertically upwards from the old position of fifty.

New defence would be 91% of the old, plus 9% of what would have resulted in the opponent kicking what they did.
In this case they kicked exactly what was expected, so the number doesn't change. If it did, the team would have moved that on the horizontal axis. Each week every team moves the combination of this horizontal and vertical movement.

So Winner team is now at 54.5 offence and 50 defence.

 

[Chunderboy]

Yes mate.

*Great post that explains it perfectly so I can understand*

Thanks a lot! I assume because you say "we use" that you work on it with Final Siren?

I was pretty sure I had the general concept, but thanks for that further explanation, I know you've probably already explained it a few times but I really appreciate you doing it again. One of my favourite threads by far!

 

[Wandering Hawk]

Let's now do the loser team.

It was on 50 offence and 50 defence.

It was expected to kick 85 and did exactly that. So we know the offence will not change. When we take 91% of the old offence and add 9% of what offence it needed to kick what it did, we get the same value, an offence of fifty. So there is no change for the loser team vertically.

Now let's look at loser's new defence. It is calculated as 91% of the old defence of fifty, plus 9% of what the defence would need to be to expect the opponent to kick the 170 they did. That's 25, so we wind up with 91% of 50 plus 9% of 25, which is 47.75 as the new defence.

So loser drops 2.25 defence, from fifty to 47.75, and moves left 2.25 on the horizontal axis.


So winner moved from 50,50 to 54.5,50 and loser moved from 50,50 to 50,47.75

Note my example was simple as they only each moved in one value, as loser scored exactly the 85 expected, but in most games both teams move on both axis.

 

[Wandering Hawk]

Nah mate just an inclusive use of we. You me and Final Siren all use this value...

 

[Sidebottom2Beams]

No score predictions this week
I love beating the squiggle while other teams dissappoint.

 

[Wandering Hawk]

Hey any chance of getting the current team Off/Def values as a table as well as a graphic?
I'd write up the predictions for y'all, but I would just be estimating values off the graph.
Or are they already somewhere and I'm just a full *?

 

[Chunderboy]

Nah mate just an inclusive use of we. You me and Final Siren all use this value...

Ahh okay, well thanks either way