Smartgambler
Pro-Punter

Go Back   OZmium Sports Betting and Horse Racing Forums > Public Forums > General Topics
User Name
Password
Register FAQ Search Today's Posts Mark all topics as read

To advertise on these
forums, e-mail us.

Reply
 
Thread Tools Search this Thread Display Modes
  #1  
Old 14th May 2016, 05:40 PM
UselessBettor UselessBettor is offline
Member
 
Join Date: Sep 2011
Posts: 3,099
Default Swarm intelligence in Kentuky Derby ... totally wrong conclusions

http://www.news.com.au/technology/i...65395637a62840c

These guys have no idea at all what they are talking about.

They apparently picked the first 4 horses in the Kentucky Derby by talking to 20 people.

Check the results here :

https://en.wikipedia.org/wiki/2016_Kentucky_Derby

Look at the odds. Hmmm the top 4 horses came in the same order as their chance. 3/1, 8/1, 10/1, 10/1.

The odds available on a horse are the results of swarm intelligence. More than 20 people putting lots of money on an outcome generally means the odds are pretty close to spot on the chance. I love how one result makes this a success.

They say it will change gambling ... lol ... they don't even understand the basics.
Reply With Quote
  #2  
Old 15th May 2016, 07:47 PM
walkermac walkermac is offline
Member
 
Join Date: Nov 2013
Posts: 901
Default

Quote:
Originally Posted by UselessBettor
These guys have no idea at all what they are talking about.

They say it will change gambling ... lol ... they don't even understand the basics.
The article reports that, but also says that the CEO "was “speechless” when the superfecta proved correct". They don't seem to give two hoots about gambling, instead using it as a means to excite interest in what is potentially a very useful tool.

From their own blog at http://unanimous.ai/unu-superfecta-11k/:
Quote:
So, when Hope Reese, a reporter for TechRepublic and the Atlantic, challenged Unanimous A.I. to use UNU to predict the winners of the Kentucky Derby, “we were reluctant to take on this challenge,” says David Baltaxe, Chief Information Officer at Unanimous. “Nobody here knows anything about horse racing, and it’s notorious for being highly unpredictable. Still, UNU surprises us again and again, so we recruited a swarm of volunteers through an online ad. The whole thing took 20 minutes.”

During an initial 10-minute session, the group used UNU to answer questions as a unified Swarm Intelligence, narrowing the field of 20 horses down to four winners. The swarm was then asked to order the four winners into Win, Place, Show, and Fourth. Then, a week later the Kentucky Derby announced the post positions of the horses, which impacts the potential outcome. So, the Swarm Intelligence was convened again, and asked if any changes should be made. One of the four picks was replaced by an alternate. This process took another 10 minutes.

Below is a replay that shows the Swarm Intelligence making one of the many decisions required to predict the full outcome of the race:
The replay referenced above was "who will be the winner?" and was completed on 29th April (our time). It isn't noted what the other questions were but, given Dave's proclamation, I doubt they were very advanced: likely 'will this horse finish top 4' to whittle down the field, then 'who gets first?', 'second?', so on...

The odds on the 27th April (their time) had Nyquist as favourite (+300, I don't know what the units mean...), Exaggerator (+800) even with Mohaymen (+800) and Gun Runner (+1000). That is, over a week prior to the actual race and in an order contrary to the betting market at the time, they predicted the final market odds order. In this instance.

The reason they were challenged by a journalist is that they've had some measure of success in the past: NFL playoff predictions (beating 12 of 13 ESPN experts), predicting 11/15 Oscar category winners (beating the New York Times and other luminaries), 13/19 in Superbowl betting (in markets as diverse as 'What colour Gatorade will be dumped on the winning coach?', beating Vegas with a 35% ROI). The last one was interesting, in that they challenged individual bettors to do better. Of the 469 entrants, only 4 had a better return. So over 99% of entrants would've been better off following the swarm predictions (http://unanimous.ai/swarm-ai-takes-...ses-and-wins-2/). They've done this challenge a couple of times, and with similar results.

With regards to the Kentucky Derby, apparently none of the 19 individual swarm members picked the correct order, nor did an unspecified number of "experts" at SBNation. Despite the finishers being in final market order.

A little more detail from their research paper (https://mitpress.mit.edu/sites/defa...027-5-ch117.pdf):
Quote:
For example, when predicting the 2015 Academy Awards,
we polled 48 individuals with a written survey, asking them to
predict the top 15 award categories. Using the most popular
predictions to represent “the wisdom of the crowd”, the group
collectively achieved 6 correct predictions for the top 15
award categories (40% success). This was our baseline
dataset, the low success rate reflecting the fact that this group
of users had no special knowledge about movies.
To test swarming, we then selected a 7 person sub-group of
the full population and asked them make the same predictions,
but now as a unified dynamic system. The 7 individuals were
typical performers on the written poll, ensuring equity with
the full 48 person population. Each of the 7 individuals were
networked over standard internet connections to a central
server from different remote locations.
Working as a unified swarm, the group of 7 individuals
achieved 11 correct predictions for the top 15 award
categories (73% success). In other words, a sub-group that
was only 15% the size of the full population had a success rate
that was nearly double. We believe this is a highly promising
result and speaks to the potential for harnessing the wisdom of
social groups through real-time swarming.
How is this useful for gambling? A much smaller population may be able to predict market consensus far earlier than the market itself resolves. The quote above indicates that swarm intelligence may also provide more accurate predictions than a poll (which is essentially what the market is: where money is votes).

Let's give it a go! You can set up your own swarm on their site: http://go.unu.ai/lobby The Doomben Cup is on Saturday. Set it up for Wednesday night? (Obviously people need to be doing it at the same time, that's how it works). Hopefully get a few people post-form study and see how it all works.
Reply With Quote
  #3  
Old 15th May 2016, 08:18 PM
Shaun Shaun is offline
Member
 
Join Date: Jan 1970
Location: Western Australia
Posts: 5,176
Default

First intelligent post in months, thanks UB some interesting stuff in there.
__________________
"If the corporates are treating you poorly , just go elsewhere."
"If they need you , they will soon find out."
"If you need them , you will soon find out."
--moeee
_______________________________________________
Reply With Quote
  #4  
Old 17th May 2016, 10:37 AM
evajb001 evajb001 is offline
Member
 
Join Date: Oct 2009
Posts: 1,224
Default

walkermac in light of the other thread being closed do you wish to email me to discuss my ideas further (if you had a chance to read them). If not all good just thought i'd check or for any other that had a chance to read my post and want to explore further.

In fact if people could email me websites that offer freely available ratings in a format that is relatively easy to scrape into excel that would be appreciated.

jbevans at internode dot on dot net
Reply With Quote
  #5  
Old 17th May 2016, 02:18 PM
walkermac walkermac is offline
Member
 
Join Date: Nov 2013
Posts: 901
Default

Quote:
Originally Posted by evajb001
walkermac in light of the other thread being closed do you wish to email me to discuss my ideas further (if you had a chance to read them). If not all good just thought i'd check or for any other that had a chance to read my post and want to explore further.

In fact if people could email me websites that offer freely available ratings in a format that is relatively easy to scrape into excel that would be appreciated.

jbevans at internode dot on dot net
UB's thread here was about a specific instance of it being used for the Kentucky Derby, whereas I wanted to discuss how we could use it as a selection method. What questions would we need to pose? Can we co-ordinate users to be online at the same time? Is it practical? Could we build more data than the one result reported? Eventually perhaps: can humans be replaced by rating methodologies that respond to/are triggered by the output of other rating methodologies to form a new system that the methods respond to in turn, and in turn, and in turn...until they convene on a selection? All of which was made clear in the first post.

JB, from what I recall of your post: no.




Elucidating a little more: what that other site you referred to does is not swarm intelligence, it's polling. The difference between the two is collaborative choice vs choice in isolation. Many choices in isolation perform OK (this is the Wisdom of Crowds), but many choosing collaboratively seems to perform better (depending on the application and design of the process).

If you've read the book The Wisdom of Crowds, you might recall the example regarding fairgoers guessing the weight of a cow: no-one got it right, but if you took the average of all the guesses - the ones that were way too high cancelled out the ones that were way too low - and they got the exact weight of the cow. A group ran the experiment again recently (this time only using a 2D photo and noting that respondents were also likely less knowledgeable regarding livestock, in this day and age). 17000 people responded and the average of all their guesses was within 5% of the actual cow's weight. They then ran the same experiment with a swarm and the swarm's answer was within 8%. (http://unanimous.ai/crowds-vs-swarms-which-is-smarter/)

The swarm did worse, so why am I using it as an example? It took a week to collate 17000 people's answers. The swarm took a few minutes and comprised 49 people. If you look even closer, the swarm's accuracy was further compromised by the design of the question. You can watch the video of the swarm operating via the link above. They first agree on a weight range, then agree on an exact weight in that range. The swarm's answer was the highest possible weight in that range. Were the range wider, they likely would have got an even closer answer.

(The ranges themselves were overlapping, which possibly divided people who agreed Penelope weighed the same, but didn't agree in the way they might be wrong. It also doesn't note whether the orientation of answers are the same for each user: humans can prefer left over right - e.g. how many right to left scrolling video games are there? - and top to bottom - eyes look top-left on a page/website first and the result there can anchor their later thoughts: https://en.wikipedia.org/wiki/Anchoring).

If you look at the histogram for the 17000 guesses of the cow's weight you'll see long tails on their distribution. Imagine you said to one of the few people on that tail who guessed a cow weighed 3000lbs: "Hey buddy, thanks for playing, but if I told you that everyone else was guessing substantially less than you, would you like to change your guess?". Unless they were particularly sure or obstinate they would change their guess to a lower one, more in keeping with their fellows. If everyone had the opportunity to continually see what their fellows guessed and altered their own estimate as a consequence, you get a swarm. In addition, you likely get a more accurate answer as instead of applying the Wisdom of Crowds to a range of 0 to infinity, you're applying the Wisdom of Crowds to a range from 1000 to 2000lbs, as the consensus resolves, for example.

The horse betting market isn't quite a swarm as it's comprised of legacy opinions (bets from hours/days/weeks before), there's no option to change a guess as the market resolves (except by throwing further money into it), is not homogeneous (some people bet thousands on their opinion, others 50c), is not free from central control (there is bookmaker's over-round, anchoring through fixed prices, laying off bets on other books, commission, etc) and people with strong opinions can choose not to participate in the market at all (when the odds aren't to their satisfaction). It does utilise the Wisdom of Crowds however and is likely why it's the most accurate, long-term measure (asides from a couple of biases).

Similarly, your ratings of ratings application is a different beast also. Each of the ratings would need to change depending on the selection and strength of confidence of all the other ratings (and the omission of horses from potential selection as the swarm narrows down to possible answers). With a static number like that from tatts, skyform, or what have you, it would be difficult to do.

Perhaps it needn't be as complicated as that:

You have an "ant" that pulls the "path" of final selection towards the horse with the highest API. The ant can determine which horse that is (of those remaining in selection) and the strength/confidence it pulls the path to this result: the historical average of 21.1% (though impact values would likely be more accurate).

You have another ant pulling in the direction of the horse with the highest winning percentage, with 18.6% of its strength. An ant pulling in the direction of highest tatts rater. An ant pulling in the direction of runners to winners percentage by sire/damsire. An ant pulling in the direction of the highest speed rating. An ant pulling in the direction of barrier stats. An ant pulling in the direction of the horse whose colour indicates its the likeliest chance. Any number of ants using every thinkable kind of determination. Hundreds or thousands of "ants" pulling the "path" to their choice as final selection. As the path moves away from particular horses, each ant omits that horse from the possible choices and recalculates the direction it's pulling in.

Either the path stays in a kind of stasis (with no one horse resolving quickly enough as a consensus selection - the UNU tool applies a 60 second limit to human swarm choices and if an answer cannot be settled upon in that time, calls it 'brain freeze' - in which case the race is omitted from betting) or the path moves to a final selection. (Alternatively they could be an equal chance).

The price could potentially be determined by how many iterations were necessary before the selection was highlighted (or reached stasis, in the case of equal chances). (That is, run it over test data. Track how "long" the final selection takes to be resolved. Group those that took a similar amount of time/iterations and then compare to the actual strike rate to determine a price).

To price the whole field, remove the existing pick/s from the selections and run the method again.

Post moderated. Comments regarding our Forum moderating efforts removed.

Last edited by Moderator 4 : 17th May 2016 at 07:59 PM.
Reply With Quote
  #6  
Old 17th May 2016, 02:50 PM
evajb001 evajb001 is offline
Member
 
Join Date: Oct 2009
Posts: 1,224
Default

Ah I see so essentially having a conversation around the selections means you can bring outliers into line a bit or persuade some peoples decisions by having coherent discussion.

For instance someone may have looked at a race and say I pick number 1 because it has a great speed profile, good barier and has won in similar class before. Person number 2 comes along and suggests that whilst all of that may be true, horse number 3 also has those same characteristics but a better jockey and its recent form is more supportive of a good run today. Person 1 may agree and adjust there selection to horse number 3 - I take it that is essentially what you're saying?

So basically a 'Rating of Ratings' approach could have a similar outcome as the swarm approach its just your taking the discussion part out of it and basically taking a mechanical approach that may need more inputs to achieve a similar outcome? i.e. your suggesting we could complete a swarm approach on the Doomben cup with say 6-10 people - does this mean we could use a "Rating of Ratings" approach with say 15 independent ratings sources and achieve a similar outcome?

I'm not suggesting we could obtain 15 freely available ratings sources but I'm just questioning whether that is another option? I might still go down the 'Rating of Ratings' path for the purpose of this exercise and maybe if we get enough people doing the swarm approach we could compare the two?

As it stands at the moment I've got 5 different ratings sources to pull in and create the "RoR" (thanks to garyf for giving me another) and we'll see if I can find 1 or 2 more to include and see how it goes.
Reply With Quote
  #7  
Old 17th May 2016, 08:51 PM
The Ocho The Ocho is offline
Member
 
Join Date: Aug 2010
Posts: 2,661
Default

Wouldn't tipsters polls do the same thing where as many tipsters polls as you can get are collated? Here in Melbourne the Herald Sun on a Saturday have a number of tipsters and their tips are already collated for you.
__________________
Never give up on a dream just because of the time it will take to accomplish it. The time will pass anyway.”

― Earl Nightingale
Reply With Quote
  #8  
Old 18th May 2016, 01:52 AM
walkermac walkermac is offline
Member
 
Join Date: Nov 2013
Posts: 901
Default

Quote:
Originally Posted by The Ocho
Wouldn't tipsters polls do the same thing where as many tipsters polls as you can get are collated? Here in Melbourne the Herald Sun on a Saturday have a number of tipsters and their tips are already collated for you.
No, it's not the same thing. A poll is a collection of opinions which are blind to each other. A swarm is a system where the collection and strength of opinions influence each other until they reach consensus.

Better description from a magazine article, (UNU is the name of the tool they used):
Quote:
Rather than voting in a static poll, UNU users control a virtual magnet that they use to pull a puck-like indicator toward their desired answer — like the Ouija board, but without the “dark magic.” The puck can only fall on one answer, and the group has 60 seconds to make a choice.

“UNU provides a continuous feedback loop of the group’s preference for a choice, as well as its conviction,” says Rosenberg. “People are adjusting their levels of conviction based on the completeness of their own knowledge on the subject.”

Users can see the speed and direction the puck is heading in real-time. That forces members of the group to constantly reevaluate their first choice. When the puck races quickly toward one option, people who aren’t as confident may join the ride and ditch their choice in favor of the group. Or, the tug of the group is too strong for one stubborn participant to overcome. In that way, the group reaches a decision optimized through group thinking.
You can play around with UNI at http://unu.ai/ (Chrome browser required). Even if there's no-one else on at the time you'll get an idea of how it works

Quote:
Originally Posted by evajb001
Ah I see so essentially having a conversation around the selections means you can bring outliers into line a bit or persuade some peoples decisions by having coherent discussion.
Close, but without the discussion. All that is indicated by an individual member of the swarm is their preference and their strength of conviction. If you add further bells and whistles the system is liable to suffer from problems like groupthink (http://www.ablongman.com/html/mindm...2/social10.html) or by paying greater attention to people who express themselves better, or to those who wear a labcoat, etc.

Quote:
Originally Posted by evajb001
So basically a 'Rating of Ratings' approach could have a similar outcome as the swarm approach its just your taking the discussion part out of it and basically taking a mechanical approach that may need more inputs to achieve a similar outcome?
It could, but the ratings would need to be able to express how sure they are in their selection, and would need to be able to be influenced by the output of all the other ratings.


Quote:
Originally Posted by evajb001
i.e. your suggesting we could complete a swarm approach on the Doomben cup with say 6-10 people - does this mean we could use a "Rating of Ratings" approach with say 15 independent ratings sources and achieve a similar outcome?

I'm not suggesting we could obtain 15 freely available ratings sources but I'm just questioning whether that is another option?
Unless you have the method of how the ratings are determined, I don't think it will work (in this application). At each movement of the group's consensus towards one selection or another, each rating needs to be recalculated to determine if its preferred selection has now changed, and its new confidence in that selection. Without that feedback loop then you're just doing a poll of the rating services. If you're scraping a static figure for the rating, you won't be able to perform those calculations.

Though if you've got a database you could make tens of simple ratings for yourself, easily enough. Every comparable form statistic is a simple rating and you can determine the strength of confidence to have in that rating. Then implement something like UNU (see pic above): each rating at that moment produces a direction (towards its preferred choice) and a velocity in that direction (its confidence). Physics determines how the consensus/puck moves, after taking all influences into account. "One second" later we determine whether any selections are out of the running (i.e. the puck is moving in the opposite direction). Each rating recalculates with the new set of possible selections. Repeat until the puck arrives at the consensus selection, or is just hovering about/moving very slowly between selections (indicating they're even chances).

...I'm not sure that would work either, but it's a closer approach to swarm intelligence. I think.

The difficulty would be that the arrangement of the horses on the "board" is arbitrary. Just because the puck is moving towards one of them doesn't mean the horse on the opposite side of the board to it is the worst of those remaining. I think human participants would handle this element much better....
Reply With Quote
  #9  
Old 18th May 2016, 08:27 AM
evajb001 evajb001 is offline
Member
 
Join Date: Oct 2009
Posts: 1,224
Default

So for instance lets say you have past data for 5 ratings sources and today's race is 1100m and has prizemoney of $20,000. Could you break the data of those 5 ratings sources down to find the best/worst performers in say these areas:

<1249m race and <$20001 Pmoney
<1249m race and <$50001 Pmoney
<1249m race and >$50001 Pmoney
<1849m race and <$20001 Pmoney
<1849m race and <$50001 Pmoney
<1849m race and >$50001 Pmoney
>1849m race and <$20001 Pmoney
>1849m race and <$50001 Pmoney
>1849m race and >$50001 Pmoney

That way you find where the race your doing the 'Rating of Ratings' approach on fits and work out how each of the 5 ratings perform in that race band. Adjust the weightings to give more weighting to the best performers and less to the worst performers (adjusting for confidence) and that brings you reasonably close to having a mechanical swarm approach with 5 sources yeh?
Reply With Quote
  #10  
Old 18th May 2016, 02:24 PM
Mark Mark is offline
Member
 
Join Date: Jan 1970
Location: Qld
Posts: 2,012
Question

And after all that, in 99% of cases you'll end up with the favourites in market order. So what's the point?
Reply With Quote
Reply


Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

vB code is On
Smilies are On
[IMG] code is On
HTML code is On
Forum Jump



All times are GMT +10. The time now is 03:30 AM.


Powered by: vBulletin Version 3.0.3
Copyright ©2000 - 2018, Jelsoft Enterprises Ltd.
©2008 OZmium Pty. Ltd. All rights reserved . ACN 091184655