Kelly betting live test - is it a madman's goal as some quants around here suggest?

[brett128]

brett128, a query here, because I've been delving deeper into my methods, etc... researching... re-inventing the wheel...

Given your comment above, if I develop an out of sample method with a MAE of 29.0 points... just under the market, would that be considered a profitable model?

If you have an MAE half a point lower than the MAE of the AFL closing line, you should be able to achieve 57% - 60% ATS long term no worries at all. However be wary of 'out of sample' testing. You should also employ such methods as cross validation in order to produce more robust out of sample results.

Personally, I've found nothing ever gives the same result in actual betting, you always have a further drop from your out of sample results. For instance, out of sample results on all our model trains do ~60%, which has translated into holding ~57% throughout the course of this season so far (356 plays now).

 

[Thedeftouch]

You can delve as deep as you want to. If you want to limit yourself to scratching the surface, that's your decision. When the utility of that method becomes incorporated into the market and your win % drops off, you're back to square one. I'd rather prefer to be so far ahead in my methodology(ies), that we won't have to worry about this for a long time. Hence the reason why we devoted some winnings to getting me overseas for some cutting edge data mining tuition.

Are you getting laid over there? Do you call it thigh mining?

Seriously though, best of luck.

 

[Rourke]

brett128, a query here, because I've been delving deeper into my methods, etc... researching... re-inventing the wheel...

Given your comment above, if I develop an out of sample method with a MAE of 29.0 points... just under the market, would that be considered a profitable model?

Interestingly this year I have the MAE for bookies' lines at about 25.5, and my model is also at 25.5. Standard deviation is around 32.5. I used to think that under ~34 points for SD was impossible, due to the natural variation just in scoring shots, and was very proud of a model that demonstrated 35.1 SD (my previous best had been closer to 37; bookies used to assume over 38). Could there be a process that is making matches more predictable? I doubt it - almost certainly just a bizarrely predictable year.

 

[lamaros]

You can delve as deep as you want to. If you want to limit yourself to scratching the surface, that's your decision. When the utility of that method becomes incorporated into the market and your win % drops off, you're back to square one. I'd rather prefer to be so far ahead in my methodology(ies), that we won't have to worry about this for a long time. Hence the reason why we devoted some winnings to getting me overseas for some cutting edge data mining tuition.

You are creating a model for 22 guys playing 22 other guys. You can call it what you want, but you are still modeling the game. The game is the basis, a game that is made up of incomprehensibly infinite factors. You are attempting to create a model that explicitly weights several factors over others, but that doesn't mean you get to ignore the game.

The human mind can intuitively and subjectively 'manage' far more factors than any model humans are capable of creating. If that mind is also well calibrated it can also be far more successful.

You are attempting to create a model to beat another model (that of the bookies) - which is somewhat different in some ways - but you are both still attempting to model a game. A game you have to respect and constantly observe for any changes that might necessitate changes in modeling.

Also you are an idiot.

 

[brett128]

The human mind can intuitively and subjectively 'manage' far more factors than any model humans are capable of creating. If that mind is also well calibrated it can also be far more successful.

That rivals some of the most moronic things I've read on this board.

So what, Google uses a bunch of human savants to manage petabyte sized data warehouses to generate data insights? Think about what you're saying and how stupid it sounds.

 

[lamaros]

That rivals some of the most moronic things I've read on this board.

So what, Google uses a bunch of human savants to manage petabyte sized data warehouses to generate data insights? Think about what you're saying and how stupid it sounds.

Computers cannot play Go at a professional level. Go is infinitely less complex than a game of AFL. Think about how stupid the things you are saying are.

Or, more likely, get someone smarter than you to explain it to you so you can then come on here and pass yourself off as a genius trying to browbeat children with it.

 

[d3ft]

That rivals some of the most moronic things I've read on this board.

So what, Google uses a bunch of human savants to manage petabyte sized data warehouses to generate data insights? Think about what you're saying and how stupid it sounds.

Computers identify patterns - humans identify insights.

I am impressed by your passion and commitment Brett, but come on dude, you're predicting sport - it's not exactly a noble pursuit. If you really cared about statistical modeling, you'd be predicting cyclones or something that might actually benefit someone other than yourself.

 

[Duritz]

If you have an MAE half a point lower than the MAE of the AFL closing line, you should be able to achieve 57% - 60% ATS long term no worries at all. However be wary of 'out of sample' testing. You should also employ such methods as cross validation in order to produce more robust out of sample results.

Personally, I've found nothing ever gives the same result in actual betting, you always have a further drop from your out of sample results. For instance, out of sample results on all our model trains do ~60%, which has translated into holding ~57% throughout the course of this season so far (356 plays now).

Yeah, have experienced that kind of thing in racing... no matter how "clean" the data, reality never mirrors your research.

Question when you're next online: is a comparison of MAE the best way to test your model, or something else?

 

[Spads]

You concept of reality expressed as through streaks is blatantly wrong, as some people have already expressed. One thing I will say though is that I think you have a Bayesian logic without even knowing it (look it up, not a bad thing at all).

I don't know. There are entire branches of mathematics devoted to random processes that are preferential towards streaks of events. In some systems, streaks or clusters are even proven to be inevitable. This is a fundamental part of telecommunications system theory, queueing theory, Markov Modulated Poisson Processes, etc.

 

[brett128]

Computers identify patterns - humans identify insights.

I am impressed by your passion and commitment Brett, but come on dude, you're predicting sport - it's not exactly a noble pursuit. If you really cared about statistical modeling, you'd be predicting cyclones or something that might actually benefit someone other than yourself.

Huh? What's a 'noble pursuit' got to do with statistical modelling? So you think I shouldn't apply the best quantitative techniques because sports betting is 'not a noble' pursuit?

Maybe leave your personal ethics at the door when talking about these fields, or maybe preach your noble pursuits to the financial sector as well, whose quantitative abilities are designed to hoover money from the common man on a gargantuan scale.

 

[brett128]

Yeah, have experienced that kind of thing in racing... no matter how "clean" the data, reality never mirrors your research.

Question when you're next online: is a comparison of MAE the best way to test your model, or something else?

There's no best way. Every way has pros and cons you need to weigh up.

MAE is one method if your predicted output is a metric variable (which you have), forward testing as if you were betting forward in time is another. Cross validation gives you big samples, but is less relevant to what is going on now in the market/sport because your taking training samples that will compose of games happening after the block of games in the testing sample.

 

[Rourke]

I don't know. There are entire branches of mathematics devoted to random processes that are preferential towards streaks of events. In some systems, streaks or clusters are even proven to be inevitable. This is a fundamental part of telecommunications system theory, queueing theory, Markov Modulated Poisson Processes, etc.

These ought to show up in analysis of the autocorrelations of your model's process vs the data. No doubt that there are reasons for some clusters, but 'winning streaks' and 'losing streaks' in betting should be given less conscious thought than our primitive pattern-detecting brain wants, since the cause is mostly random. I could start on 'momentum' but I won't ...

 

[Duritz]

These ought to show up in analysis of the autocorrelations of your model's process vs the data. No doubt that there are reasons for some clusters, but 'winning streaks' and 'losing streaks' in betting should be given less conscious thought than our primitive pattern-detecting brain wants, since the cause is mostly random. I could start on 'momentum' but I won't ...

Momentum... tell the Norf players momentum doesn't exist!!

RE streaks, I do prefer to think of them as clusters, and I won't rewrite what I already said, but my overall theory is given we can't quantify every variable (not even close in horse racing) then it's entirely plausible (IMO) to suggest that things can come and go (like weather) beyond our control that can cause clusters, both good and bad.

If we could quantify every variable, like in a coin toss, then yes I'd agree there's no such thing. But I'd argue in assessing probabilities for racing or sport, we can never, ever quantify every variable.

Unless we're God.

Which I'm pretty sure no-one on Bigfooty is.

 

[Duritz]

There's no best way. Every way has pros and cons you need to weigh up.

MAE is one method if your predicted output is a metric variable (which you have), forward testing as if you were betting forward in time is another. Cross validation gives you big samples, but is less relevant to what is going on now in the market/sport because your taking training samples that will compose of games happening after the block of games in the testing sample.

Over a decent sample, what's your MAE at, if I may ask?

 

[Kynge of Begrem]

. Hence the reason why we devoted some winnings to getting me overseas for some cutting edge data mining tuition.

Dude, when you have posted bets here you have shown that you lost 87 per cent of your money, you have no winnings, the only way you are getting overseas is by swimming.

You are making this stuff up.

 

[Thedeftouch]

So..... will there be any more betting?

Brett any updates from your overseas junket?

 

[Duritz]

So..... will there be any more betting?

Brett any updates from your overseas junket?

There MIGHT be... though it may be modified.

 

[brett128]

Over a decent sample, what's your MAE at, if I may ask?

I don't build my AFL model output in points, but in the direct probability of covering ATS. I do have points based sub-models, but they all get converted into a win % so I can stitch it together with everything else.

 

[brett128]

So..... will there be any more betting?

Brett any updates from your overseas junket?

Was only in Europe for 10 days, which is a pity. Am currently in Amsterdam now, about to hop onto a plane to Guangzhou, on my way back to Australia.

Only took 3 days off for myself as a long weekend down to Austria, caught the train down from Dortmund, Germany. Went to a little town called Voralsberg, about 70 km east of Lindau for a traditional beer/music festival, kind of like Oktoberfest but with no tourists. Caught up with some Aussie mates that work there, and their Austrian friends. Had one of the best weekends of my life I reckon, I wore lederhosen, drank copius amounts of beer and was surrounded by women that wore the traditional drindl get up.

 

[Duritz]

Was only in Europe for 10 days, which is a pity. Am currently in Amsterdam now, about to hop onto a plane to Guangzhou, on my way back to Australia.

Only took 3 days off for myself as a long weekend down to Austria, caught the train down from Dortmund, Germany. Went to a little town called Voralsberg, about 70 km east of Lindau for a traditional beer/music festival, kind of like Oktoberfest but with no tourists. Caught up with some Aussie mates that work there, and their Austrian friends. Had one of the best weekends of my life I reckon, I wore lederhosen, drank copius amounts of beer and was surrounded by women that wore the traditional drindl get up.

If you didn't drink Austrian Riesling while there, and Gewurtztraminer, do next time. Brilliant and relatively unknown wines from Austria. WW2 had a lot to do with that.

 

[Kynge of Begrem]

Was only in Europe for 10 days, which is a pity. Am currently in Amsterdam now, about to hop onto a plane to Guangzhou, on my way back to Australia.

Only took 3 days off for myself as a long weekend down to Austria, caught the train down from Dortmund, Germany. Went to a little town called Voralsberg, about 70 km east of Lindau for a traditional beer/music festival, kind of like Oktoberfest but with no tourists. Caught up with some Aussie mates that work there, and their Austrian friends. Had one of the best weekends of my life I reckon, I wore lederhosen, drank copius amounts of beer and was surrounded by women that wore the traditional drindl get up.

Of course you did, of course you did

 

[KardiniaPark]

I think he may be the biggest douchebag ever, that was truly amazing.

 

[Kynge of Begrem]

I think he may be the biggest douchebag ever, that was truly amazing.

Sorry to disagree with you KP, but the biggest douchebag ever was a guy I met at the annual donkey dick convention I go to every year, you can only go if you have an unfeasably large penis.

Please believe that, despite of all the photographic and medical evidence, witness statements etc that my manhood has shrunk 87 per cent (evidence which I myself have posted on my official penis thread), I am in fact hung like a donkey. I will keep reappearing to ensure that people know this about me.

Really it has gotten huge since I stopped posting, I am finalising a special treatment plan that you can all use - if you are interested in buying it from me

 

[Duritz]

So, the old index stands at 39.7 ATM. That ain't a good index.

Do I dare... do I... dare?

YES I DOOOOOO!!!!!!

Carlton, H2H, 10 index points (1k) on them at $2.37 with topsport!!!

BANG!!!

 

[Duritz]

I'm BACK!!!