Certified Legendary Thread Race for the flag, in squiggly lines

[lamby]

It's pretty easy to tell how effective most of the models floating around are, because their performance is documented over several years. They generally go at around 68-74%, depending on the year. If you track your own tipping, you can see whether you beat them or not.

The kind of perfect model you describe, which uses dozens of different inputs, doesn't sound very good to me, because it would inevitably introduce more noise than signal. This is the overfitting problem, where if you give a model a ton of different factors, it will fit the historical data really well, but can't predict anything, because all it does is find a bunch of coincidences. For example, I bet you could find one particular seat at the MCG where almost every game in 2016 was won by the team whose supporter sat there. That fits the data well but has no predictive power.

The smart way to use a model, in my opinion, is as a tool rather than a Magic Eight-Ball. Because while computer models are better than most punters, they're not better than good punters. They're really best at aggregating large amounts of data that humans have trouble remembering all at once. So you don't want to take their predictions as gospel, but if you can understand what they're measuring and why they're reaching the conclusions they are, that's a good input for the model you're running inside your human brain.

Also, you know, when you just haven't been paying attention to the footy recently, then that's a good time to snag their tips.

I heard a guy talk a a big data conference where they did analysis on e-harmony results. They take hundreds on data points that people enter about themselves (height, weight, income, movie likes etc) and tried to make the perfect matching model. They wanted to know which data points were 'signal' and which were 'noise'.

What they found was that it was almost all noise. What they found was that there were really only 2 data points that were better than random picks for a successful matching of a couple - 'time of day of entering the profile' with 'type of camera used to take profile picture'!

So adding more information into a model does not always improve it - it just adds more noise!

 

[Seeds]

Ok my previous posts were perhaps a bit too harsh. I'm not saying there are currently better models around as far as I know there aren't. However, the big data revolution is just around the corner and it's going to produce models that make these ones completely obsolete. Don't pray at the foot of this model as it's about to be knocked on its head unless ofcourse Mr Squiggle taps into the same data revolution.

I don't gamble on football because the models we have aren't superior enough to beat the bookies and the risk premiums they apply on the punter to ensure they make a profit over the long term. Would much rather gamble on stocks where there is no risk premium to gamble against.

 

[Seeds]

I heard a guy talk a a big data conference where they did analysis on e-harmony results. They take hundreds on data points that people enter about themselves (height, weight, income, movie likes etc) and tried to make the perfect matching model. They wanted to know which data points were 'signal' and which were 'noise'.

What they found was that it was almost all noise. What they found was that there were really only 2 data points that were better than random picks for a successful matching of a couple - 'time of day of entering the profile' with 'type of camera used to take profile picture'!

So adding more information into a model does not always improve it - it just adds more noise!

This is true. It has to be the right information. But ultimately the world is just one giant model of data flows. Theoretically the information in the world exists to predict with 100 percent accuracy. The question is how much of that information is attainable.

 

[Seeds]

But remember, The Squiggle was not created to be the best tipper or the best model. It was created to be a visual indicator of where teams sit both comparatively to eachother and over time.

That makes it interesting but not useful.

 

[gavaniacono]

This is true. It has to be the right information. But ultimately the world is just one giant model of data flows. Theoretically the information in the world exists to predict with 100 percent accuracy. The question is how much of that information is attainable.

Thats only true if the world is both finite, and static. if the world is one of infinite or dynamic where conditions and therefore effects change then no, 100 predictive accuracy is not possible.

 

[SJ]

This is true. It has to be the right information. But ultimately the world is just one giant model of data flows. Theoretically the information in the world exists to predict with 100 percent accuracy. The question is how much of that information is attainable.

From what I understand, that is not true at the quantum level where the 'natural laws' (as we know them) break down to uncertainty and a probabilistic reality.

 

[gavaniacono]

From what I understand, that is not true at the quantum level where the 'natural laws' (as we know them) break down to uncertainty and a probabilistic reality.

Correct weight Randwick Quantum theory as a complex probability theory. 100 % accuracy - nope. Useful - very much so.

 

[Dr Tigris]

Ok my previous posts were perhaps a bit too harsh. I'm not saying there are currently better models around as far as I know there aren't. However, the big data revolution is just around the corner and it's going to produce models that make these ones completely obsolete. Don't pray at the foot of this model as it's about to be knocked on its head unless ofcourse Mr Squiggle taps into the same data revolution.

I don't gamble on football because the models we have aren't superior enough to beat the bookies and the risk premiums they apply on the punter to ensure they make a profit over the long term. Would much rather gamble on stocks where there is no risk premium to gamble against.

Common point of view. But as a researcher that uses data in his job - i.e. that's what I do - I suspect that big data just won't produce such a super model. It won't happen because of the reason Final Siren gave. All models are essentially backward looking. When you are dealing with natural processes that have invariant underlying dynamics they can reveal heaps. When humans get into the game it all falls apart. Humans innovate. Thus they change the underlying dynamics.

Put another way. Gravity is gravity and operates the same everywhere all the time. When the press was invented and 'perfected' it changed the game. Until coaches worked out how to get around it. What a model can say, at best, is that given this set of constructs, connected together in this specific way, with these theorized relationships, these inputs should lead to this pattern of outputs. If someone changes the underlying dynamics/rules by innovating the model is flawed. One thing we know footy is always innovating. Models can help, but only help.

And I am not even talking of complex systems. In a complex system, like most social systems - e.g. sports - very minor changes in initial conditions can lead to dramatically different outcomes in systems where the dynamics are known perfectly. Put another way an issue like a lucky bounce leading to a lucky goal early in the season that leads to a lucky win can lead to improved morale and belief in the game plan, that then leads more wins and so better performance. That says that luck and historical lock in can drive results in a way that no model can perfectly capture.

Big data is really useful for complex situations where there is lots of very well specified, high quality data, that provides insight in to key features of the system, so that data mining (or similar) can find insights that aren't obvious. Models might improve. But I would personally go for a simpler model that captures high level features of what is going on, and then use expert intuition to get better results in picking winners.

Studies have shown that an expert's intuition in their area of expertise is better than that of models built by experts with lots of variables. But as soon as you move from their area of expertise the models kill them. From this I gather that models are much better than a good guess, but worse than expert judgment.

 

[Russian Demon]

Correct weight Randwick Quantum theory as a complex probability theory. 100 % accuracy - nope. Useful - very much so.

Whats the quote?

All models are wrong. Some models are useful.

 

[gavaniacono]

Whats the quote?

All models are wrong. Some models are useful.

Cite global warming models for a laugh, if you need one. Quote: some models are hilarious.

 

[Seeds]

Thats only true if the world is both finite, and static. if the world is one of infinite or dynamic where conditions and therefore effects change then no, 100 predictive accuracy is not possible.

I don't believe in alternate universes as it would mean there are universes where I am sometimes wrong. The world is finite.

 

[Russian Demon]

Cite global warming models for a laugh, if you need one. Quote: some models are hilarious.

Yeah they are pretty bad. But its hardly surprising considering the number of variables combined with the relatively poor understanding of the fundamental laws that govern those variables.

Its interesting to compare it to modelling chemical processes.... where the number of variables is huge, but the understanding of how they interact and behave is much more advanced. When there is understanding, its easier to shave off the excess and make something both reasonably accurate and useable.

 

[gavaniacono]

Yeah they are pretty bad. But its hardly surprising considering the number of variables combined with the relatively poor understanding of the fundamental laws that govern those variables.

Its interesting to compare it to modelling chemical processes.... where the number of variables is huge, but the understanding of how they interact and behave is much more advanced. When there is understanding, its easier to shave off the excess and make something both reasonably accurate and useable.

Im no scientist and sometimes one needs not be to notice major influential variables not being included in a model. but ive seen plenty of very useful and indicative ones.

 

[gavaniacono]

I don't believe in alternate universes as it would mean there are universes where I am sometimes wrong. The world is finite.

Well it may well be and you may be right. However there are good arguments the other way. ive gone back and forth, currently on the infinite side.

 

[The Speaker]

It's pretty easy to tell how effective most of the models floating around are, because their performance is documented over several years. They generally go at around 68-74%, depending on the year. If you track your own tipping, you can see whether you beat them or not.

The kind of perfect model you describe, which uses dozens of different inputs, doesn't sound very good to me, because it would inevitably introduce more noise than signal. This is the overfitting problem, where if you give a model a ton of different factors, it will fit the historical data really well, but can't predict anything, because all it does is find a bunch of coincidences. For example, I bet you could find one particular seat at the MCG where almost every game in 2016 was won by the team whose supporter sat there. That fits the data well but has no predictive power.

The smart way to use a model, in my opinion, is as a tool rather than a Magic Eight-Ball. Because while computer models are better than most punters, they're not better than good punters. They're really best at aggregating large amounts of data that humans have trouble remembering all at once. So you don't want to take their predictions as gospel, but if you can understand what they're measuring and why they're reaching the conclusions they are, that's a good input for the model you're running inside your human brain.

Also, you know, when you just haven't been paying attention to the footy recently, then that's a good time to snag their tips.

This is the fundamental reason why computer models totally failed at predicting the US presidential election.

 

[LittleG]

Can we just accept that the Squiggle is great!
It's not every year we have a Bulldog level event to question our faith. Reaffirm your Squiggle faith by looking at every other squiggly year. This is currently the best football prediction model there is. It's clear, follows mathematical rules and isn't made up like power rankings.

 

[gavaniacono]

Can we just accept that the Squiggle is great!
It's not every year we have a Bulldog level event to question our faith. Reaffirm your Squiggle faith by looking at every other squiggly year. This is currently the best football prediction model there is. It's clear, follows mathematical rules and isn't made up like power rankings.

No models can predict Black Swan events. The fourth quadrant. And thank goodness for that and the Dogs.

 

[SJ]

This is the fundamental reason why computer models totally failed at predicting the US presidential election.

They didn't totally fail and it's actually hard to prove either way with a limited sample size. 538 gave Trump around 30 per cent shot, which is a reasonable chance indeed.

 

[The Speaker]

They didn't totally fail and it's actually hard to prove either way with a limited sample size. 538 gave Trump around 30 per cent shot, which is a reasonable chance indeed.

Which is the equivalent of "Because while computer models are better than most punters, they're not better than good punters." Most punters got the election wrong too...but the good ones got it spot on.

 

[LittleG]

This is the fundamental reason why computer models totally failed at predicting the US presidential election.

Computers didn't fail, the programmers did. They made so many false assumptions, that they were calculating the popular vote and not the regional vote. These are some of their terribly inept assumptions:
-Uniform national swing
-all votes are equal
-African Americans will all turn out as strongly for Hillary as for Obama
-Likely voters will vote in a uniform percentage for each candidate
-polls in all states matter
-equal waiting of polls in major cities, compared to minor cities

Trump targeted just 4 states and the 64 votes he needed to win. Hillary targeted every vote and tried to appeal to everyone.
I could write another 40+ pages on this but in conclusion, computers are only as good as their programming. Until we make AI that program themselves.

 

[Dogs_R_Us]

Should there be a fresh thread for 2017?

 

[GreyCrow]

Should there be a fresh thread for 2017?

I agree. I asked this a month or so ago. Great thread and should be in the HOF , but its getting bloated in my opinion.

 

[gavaniacono]

Computers didn't fail, the programmers did. They made so many false assumptions, that they were calculating the popular vote and not the regional vote. These are some of their terribly inept assumptions:
-Uniform national swing
-all votes are equal
-African Americans will all turn out as strongly for Hillary as for Obama
-Likely voters will vote in a uniform percentage for each candidate
-polls in all states matter
-equal waiting of polls in major cities, compared to minor cities

Trump targeted just 4 states and the 64 votes he needed to win. Hillary targeted every vote and tried to appeal to everyone.
I could write another 40+ pages on this but in conclusion, computers are only as good as their programming. Until we make AI that program themselves.

Additionally, Hillary didnt excite her base to vote generally (contrast Bill and Obama), Trump had hidden voters not picked up in polls, and contrary to opinion Trump did well with richer folk.

 

[Ressies runner]

I heard a guy talk a a big data conference where they did analysis on e-harmony results. They take hundreds on data points that people enter about themselves (height, weight, income, movie likes etc) and tried to make the perfect matching model. They wanted to know which data points were 'signal' and which were 'noise'.

What they found was that it was almost all noise. What they found was that there were really only 2 data points that were better than random picks for a successful matching of a couple - 'time of day of entering the profile' with 'type of camera used to take profile picture'!

So adding more information into a model does not always improve it - it just adds more noise!

I personally know 6 couples in long term relationships from eharmony with myself included the noise obviously works

 

[Final Siren]

Should there be a fresh thread for 2017?

Yeah there will be when I have 2017-relevant content.