Certified Legendary Thread Race for the flag, in squiggly lines

[telsor]

Carlton, Suns & Saints trying to clump together to hide behind each other (and maybe find if they can get the ultimate hidey place by going off the chart).

North & Port doing a synchronized diving routine.

 

[Randomizor]

Carlton, Suns & Saints trying to clump together to hide behind each other (and maybe find if they can get the ultimate hidey place by going off the chart).

North & Port doing a synchronized diving routine.

The squiggle taking after Lindsay Thomas for North?

 

[telsor]

The squiggle taking after Lindsay Thomas for North?

He's their new squiggle coach

 

[Chaisa]

Wow, only one win seperates 6-13. And the Giants have a very similar percentage to them as well.

Gonna be a very interesting next few rounds.

 

[jod23]

This weekend will tell us a lot about WC. Whilst we have had good wins against Port in Adelaide and GWS at home, most of our other wins have been expected. Geelong are a middle of the road team, just about where I expected us to be this season also. Home ground helps but if we can win then I might start believing we can actually finish top 4.

 

[Purple_Blood93]

Squiggle: Port, your making me look stupid.

Port: We're sor-

Squiggle: ..don't want to hear it! You'll now finish 8th and if you pull any more **** like that, you won't be making the eight.

I told you, Lord Squiggle wouldn't be pleased...

 

[NarniMagic]

A massive logjam in the middle there from Collingwood to the Bulldogs. An incredibly even season.

 

[woosha24]

West Coast briefly touched the Fremantle symbol last week.

Our new goal is to touch Hawthorn, so we're heading up.

WCE - the pedobear of the AFL.

 

[im_a_lazy_sod]

West Coast briefly touched the Fremantle symbol last week.

Our new goal is to touch Hawthorn, so we're heading up.

WCE - the pedobear of the AFL.

 

[Purple Suit]

This weekend will tell us a lot about WC. Whilst we have had good wins against Port in Adelaide and GWS at home, most of our other wins have been expected. Geelong are a middle of the road team, just about where I expected us to be this season also. Home ground helps but if we can win then I might start believing we can actually finish top 4.

It will be very interesting this week. Geelong have the tall forwards to stretch your undermanned defense. Absolutely pummelled you last year in the same game so they won't have any fear travelling over. Two totally different sides though. Beat them and I'm convinced.

 

[jod23]

It will be very interesting this week. Geelong have the tall forwards to stretch your undermanned defense. Absolutely pummelled you last year in the same game so they won't have any fear travelling over. Two totally different sides though. Beat them and I'm convinced.

Yeah I think it's a real test for us, particularly down back. Can McGovern, Hurn and Shoey hold up. We are so short on tall defenders. Plus Yeo is out suspended too.

 

[Suntory Boss Coffee]

From my understanding the Squiggle uses the last 2 years of data to make the positioning and such.
Have you ever used the last 3 years and seen if the squiggle provides more accurate predictions?

 

[KillaKanga]

From my understanding the Squiggle uses the last 2 years of data to make the positioning and such.
Have you ever used the last 3 years and seen if the squiggle provides more accurate predictions?

Most teams change a bit in 3 years. Style of play, cohesion, players in/out, form.

The Kangaroos have changed several times over 3 years

 

[Suntory Boss Coffee]

Most teams change a bit in 3 years. Style of play, cohesion, players in/out, form.

The Kangaroos have changed several times over 3 years

Yeah I know. I am more curious on whether the predictor changes or not. Becoming more accurate or less.

 

[Final Siren]

From my understanding the Squiggle uses the last 2 years of data to make the positioning and such.
Have you ever used the last 3 years and seen if the squiggle provides more accurate predictions?

I have, and no, it doesn't. See here:

Beyond 30 games or so, it seems unlikely that the extra data makes any difference, since it only affects a team's position by a small amount.

As an example, ISTATE-91:12 (the original algorithm) gets 149 tips right in 2014 under the current system, where it starts with one year's data and adds to it from there, finishing with two full seasons' data. With a rolling window, it goes like this:

Rolling window (# of games)|Correct tips \ 38-55 | 147 \ 35-37 | 148 \ 32-34|149 \ 22-31|148 \ 21|147 \ 20|149 \ 19|148 \ 16-18|147 \ 15|144 \ 14|139 \ 12-13|137 \ 11|135 \ 10|133 \ 8-9|132 \ 7|129 \ 6|126 \ 5|128 \ 4|126 \ 3|129 \ 2|124 \ 1|120 \ 0|119

 

[lusheslewis]

Fascinating stuff. Is there any sort of weighting applied to more recent results? If not, could that be easily tried with the current algorithms?

Also how does it work with 0 games? Does it just tip the home team?

 

[Billy ray]

as a gold coast supporter, and looking thru the rest of season forecast, I am really looking forward to the bye weekend.

 

[beert]

Fascinating stuff. Is there any sort of weighting applied to more recent results? If not, could that be easily tried with the current algorithms?

Also how does it work with 0 games? Does it just tip the home team?

I presume that interstate games are decided by home ground advantage and other games are tipped as draws its necessary.

 

[Final Siren]

Fascinating stuff. Is there any sort of weighting applied to more recent results? If not, could that be easily tried with the current algorithms?

Also how does it work with 0 games? Does it just tip the home team?

The "91" in ISTATE-91:12 means that each round, 91% of a team's position is dictated by previous matches and 9% by the new match. (The "12" is for 12 points of home game advantage.) This works as a weighted average like this:

Game Recency | Weighting \ 1 | 9% \ 2 | 8.2% (9% x 91%) \ 3 | 7.5% (9% x 91% x 91%) \ 4 | 6.8% \ 5 | 6.2% \ 6 | 5.6% \ 7 | 5.1% \ 8 | 4.7% \ 9 | 4.2% \ 10 | 3.9% \ 11 | 3.5% \ ... | ... \ 22 | 1.2% \ 23 | 1.1% \ 24 | 1.0% \ ... | ... \ 30 | 0.6% \ 31 | 0.5% \ ... | ...

I use these values because I just let my system run simulations of every possible combination of values (ISTATE-99:0, ISTATE-99:1...) for the last 5, 10, 15, 20, 25, and 30 years, and these ones came out best.

Since there's not always 30+ games of previous data -- like at the start of a year, when teams are reset with only the previous season's results -- the default 50/50 starting position can hold quite a lot of weight, too. If you add up all the percentages for games 1-22, you get 87.4%, meaning there's 12.6% left for the default 50/50 position. This has the effect of dragging all teams back towards 50/50 at the start of a season, then letting them spread out as each new game reduces the weighting of the default starting position.

More info: here -> Info -> Model.

When teams are rated identically -- which never happens in practice -- the squiggle will tip the home team by 0 points. It will also tip a team by 0 points if the predicted margin is less than 0.5pts, which does happen occasionally. For the purposes of tipping, you should interpret this as a tip by 1 point, since it never makes sense to tip a draw. But I've let it display "by 0 pts" so you can tell the difference between a predicted margin of 0.5-1.4 points and one of less than 0.5 points.

 

[Final Siren]

as a gold coast supporter, and looking thru the rest of season forecast, I am really looking forward to the bye weekend.

The predictor assumes the Suns get no better or worse, though. Given their current injuries, youth, and the new coach factor, you'd think they have plenty of scope to improve their form this year and not much potential to get worse. So I'd probably look at their forecast as a worst case scenario.

 

[cryptor]

The "91" in ISTATE-91:12 means that each round, 91% of a team's position is dictated by previous matches and 9% by the new match. (The "12" is for 12 points of home game advantage.) This works as a weighted average like this:

Game Recency | Weighting \ 1 | 9% \ 2 | 8.2% (9% x 91%) \ 3 | 7.5% (9% x 91% x 91%) \ 4 | 6.8% \ 5 | 6.2% \ 6 | 5.6% \ 7 | 5.1% \ 8 | 4.7% \ 9 | 4.2% \ 10 | 3.9% \ 11 | 3.5% \ ... | ... \ 22 | 1.2% \ 23 | 1.1% \ 24 | 1.0% \ ... | ... \ 30 | 0.6% \ 31 | 0.5% \ ... | ...

I use these values because I just let my system run simulations of every possible combination of values (ISTATE-99:0, ISTATE-99:1...) for the last 5, 10, 15, 20, 25, and 30 years, and these ones came out best.

Since there's not always 30+ games of previous data -- like at the start of a year, when teams are reset with only the previous season's results -- the default 50/50 starting position can hold quite a lot of weight, too. If you add up all the percentages for games 1-22, you get 87.4%, meaning there's 12.6% left for the default 50/50 position. This has the effect of dragging all teams back towards 50/50 at the start of a season, then letting them spread out as each new game reduces the weighting of the default starting position.

More info: here -> Info -> Model.

When teams are rated identically -- which never happens in practice -- the squiggle will tip the home team by 0 points. It will also tip a team by 0 points if the predicted margin is less than 0.5pts, which does happen occasionally. For the purposes of tipping, you should interpret this as a tip by 1 point, since it never makes sense to tip a draw. But I've let it display "by 0 pts" so you can tell the difference between a predicted margin of 0.5-1.4 points and one of less than 0.5 points.

Do you think there could be any improvement if the weightings of more recent games increased by a certain amount as the current season progresses?

Just to clarify, I realise that 9% is applied to the most recent game, 8.2% to the next most recent game, etc, etc as the season progresses as a rolling data set. What I mean is that as the current season progresses could the weightings for the more recent games be increased (eg. 9.2%, 8.3%, etc, etc.) to provide further relevance to the back half of the current season? With the closer we get to the GF the more weighting is applied to the most recent games and less applied to games from last season and possibly even in the opening few rounds (whatever simulations might find is the best indicator). Would be a bit complicated to implement I imagine.

 

[exe ex machina]

Look for the Blues to win out the rest of the season, claim a finals berth, and win the flag.

 

[Oneiros]

It will be very interesting this week. Geelong have the tall forwards to stretch your undermanned defense. Absolutely pummelled you last year in the same game so they won't have any fear travelling over. Two totally different sides though. Beat them and I'm convinced.

I sat through that no-goal-after-quarter-time effort at Kardinia, and sat through Geelong handing us our lowest score at home the other year as well. I cannot for the life of me understand being favourites with the bookies.

 

[Final Siren]

Do you think there could be any improvement if the weightings of more recent games increased by a certain amount as the current season progresses?

Just to clarify, I realise that 9% is applied to the most recent game, 8.2% to the next most recent game, etc, etc as the season progresses as a rolling data set. What I mean is that as the current season progresses could the weightings for the more recent games be increased (eg. 9.2%, 8.3%, etc, etc.) to provide further relevance to the back half of the current season? With the closer we get to the GF the more weighting is applied to the most recent games and less applied to games from last season and possibly even in the opening few rounds (whatever simulations might find is the best indicator). Would be a bit complicated to implement I imagine.

I have an algorithm like that called MOVER, although it works the opposite way you describe: it cares more about recent form at the start of a season, based on the idea that teams can change a lot over the off-season. So it jumps around at the start of the year before settling down.

Doing it the other way would mean the squiggle would take longer to react to teams that get significantly better or worse over a summer. The other main problem I see is that very good teams often underperform in the last few home & away rounds, like Fremantle getting creamed by St Kilda by 71 points in the last round of 2013, because they care so little about the result. You actually want to ignore that kind of thing, not weight it more heavily.

I just ran a simulation of REVERSEMOVER variations over the last 10 years and yeah, it did badly. All 5,600 possible variations performed worse than ISTATE-91:12.

But I do think it's a good concept for finals, and this is why the Grand Final tip comes from OFFDEF-75, which cares more about recent results, weighting the prelim final at 25% and the final before that at 19%.

 

[cryptor]

I have an algorithm like that called MOVER, although it works the opposite way you describe: it cares more about recent form at the start of a season, based on the idea that teams can change a lot over the off-season. So it jumps around at the start of the year before settling down.

Doing it the other way would mean the squiggle would take longer to react to teams that get significantly better or worse over a summer. The other main problem I see is that very good teams often underperform in the last few home & away rounds, like Fremantle getting creamed by St Kilda by 71 points in the last round of 2013, because they care so little about the result. You actually want to ignore that kind of thing, not weight it more heavily.

I just ran a simulation of REVERSEMOVER variations over the last 10 years and yeah, it did badly. All 5,600 possible variations performed worse than ISTATE-91:12.

But I do think it's a good concept for finals, and this is why the Grand Final tip comes from OFFDEF-75, which cares more about recent results, weighting the prelim final at 25% and the final before that at 19%.

Would it be possible for you to include these other models into a drop down selection box on the squiggle page for people to see how they differ first hand? Assuming it's not too difficult to add and you've got the time to do so of course.