Neural Networks as a means of prediction.

[merlin]

May I say a big hello to one and all. I have not been a PL member long in terms of posts etc, but I am fairly experienced in dabbling with all sorts of systems over the years in an endeavour to find something that will accurately predict matches with a fairly high percentage rate.I suppose that we all long for the perfect system(if there is one) but the next best will do for starters.:D It is with this in mind and with members with a similar view, that I introduce Neural Network Systems with a view for further discussion.Are there any NN users/supporters who dabble with these willing to share their views/ideas for mutual furtherance of such systems? Cheers. Merlin.;)

[przeszczepan]

Re: Neural Networks as a means of prediction. Far from being an expert on the subject but having read a thing or two and experimented a bit, I have come up with a conclusion that there's not much left for an artificial neural network in the field of football prediction when you have applied conventional statistical and econometrical methods. But again, I am not an expert definitely.

[spreadman]

Re: Neural Networks as a means of prediction. HI THERE MERLIN, I, too, am new to the p.l.- I think you would possibly achieve better input response if you gave more details as to what exactly 'neural networks' are. I am not afraid to admit, even though I have been betting for many years and feel that I know more than most about spreadbetting, (which can be an area of 'event determinant investment' foreign to many bettors), my knowledge of `N.N.'s` is minimal. If you have the time and patience to explain in more detail the workings in a betting context I would greatly appreciate this, as I believe you can teach on 'old dog', (me!), new tricks. Spreadman.(from one wizard to another!)

[merlin]

Re: Neural Networks as a means of prediction. Hi there Przeszczepan, hope I got the spelling right! Thanks for your initial input, a little scanty but nevertheless a reply. When you say that there is not much left in the field of NN and artificial network, what makes you think this? What sort of variables have you used? There possibly have been scores of different variables used, but we look for the solution of which ones would be the best to use in the prediction process.Have you ever used any of the NN software packages on the market, some are good, and some not so good.Currently I have three, but still search for good input and more importantly, good output!Please feel free to return here and let me know how far down the line you are with these.Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction. Hi there young Spreadman, thanks also for your initial input. I note that you are more of a spreadbetting expert, I know absolutely zero about this, so perhaps we can prop each other up on the way ahead.Neural Networks is an Artificial Intelligence which is usually a search and optimization technique program based on genetic algorithms, designed to search for the best combination of input data,the most predictable forecast horizon for any given problem, and then automatically creates and optimizes forecaster structures and control parameters by evolution and genetic search.Sounds a little long winded. It is really a program which remembers and learns from the data as and when presented.There are numerous uses within Industry etc for Neural Networks, they can also be used for Stockmarket prediction with a great degree of accuracy.The learning process starts immediately with the first set of variables fed into the program, and as these differ due to circumstance, the evolution process kicks in and remembers everything that has gone before.Networks are built up and generation after generation of data are taken into this process in an endeavour to optimize results.There are many write ups on the internet in relation to Neural Networks if you need to familiarise yourself further.Software packages differ from one to the other,there are those that cost an arm and a leg, whilst others are affordable.Personally speaking, I think at the end of the day, it is solely down to the sets of variables that are used, and just how relative they are in the prediction process.Judging the importance of these is a fine art and involves many an hour attempting to find those which you think will give the best results overall.Must dash now, but hope to discuss further.;)

[oddsfellow]

Re: Neural Networks as a means of prediction. Merlin, I do use NN's but am far from an expert. I think the saying Garbage In Garbage Out was invented for NN use. It takes careful consideration of not only what factors to include but also how to present those factors to the software. You have to remember at all times that the software views it as purely a bunch of numbers and does not know in what context you are presenting it. Peter May is the horse racing authority on this and his website is http://www.pjmracing.com/

[muppet77]

Re: Neural Networks as a means of prediction. check this "There are many artificial intelligence processes that can be used including Neural Networks, Rule Extraction, Cluster Analysis and Statistical Analysis. During the development of this site various combinations of these processes have been used. "

[muppet77]

Re: Neural Networks as a means of prediction. oh dear........ "For the English Premiership, the aiBET database has 8,511 results from Aug 25, 1984 to Mar 20, 2005. Since Oct 2004: 41 correct from 81 bets (51 % SR), level stakes returning 73.6 (-9 % RoI)."

[merlin]

Re: Neural Networks as a means of prediction. Hi there Oddsfellow. Thanks for your further input, I note also that you have used NN's and as I understand it, you still do.I don't go along with you when you say that the NN's are just presented with a bundle of numbers and therefore have no recognition as to input and output.If the data is presented in a methodical and logical manner, then logical output is not a problem.One has to determine what is worthwhile data and that which is not.I have had corners and attacking play mooted as data in the past, but just what effect and cause has this on the end product?I would certainly say that not all my own output has been useful enough to use as predictive material.However it is not all doom and gloom, as experiments including Ladbrokes odds have yeilded some surprise results lately where value bets have shown up.It's early days in this experiment,so will have to comment further on this at a later date.Can I ask you what sort of variables you are using, and maybe we can talk further on this.Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction. Hi Muppet just what can I say.You seem intent on presenting an abrasive mannerism both on Mr.Onemore's thread and now on this one.When I started this thread it was to invite constructive thoughts and ideas in the further usage and development of Neural Networks as a prediction tool.Here you are rushing in with quotes gleened from surfing the Internet that really are meaningless.Your message number 7 for example is just a straight copy and paste from another site, and offers no food for thought whatsoever.Your next post smacks of sarcasm and begins...... oh dear! Again another whole host of figures is quoted without any substance and again of destructive nature.Just what is your own experience in the usage of Neural Networks, can you present any valid discussion that may contribute to what could fast become an interesting thread?I do not know who crossed you, nor am I interested, but two wrongs do not make a right.Let's start again, I extend the olive branch and hope that you feel as if you can contribute in a worthwhile manner.:hope

[oddsfellow]

Re: Neural Networks as a means of prediction.

Merlin, I think the saying Garbage In Garbage Out was invented for NN use. It takes careful consideration of not only what factors to include but also how to present those factors to the software. You have to remember at all times that the software views it as purely a bunch of numbers and does not know in what context you are presenting it.

I obviously didn't explain this very well because i certainly didn't mean the inputs or outputs have no value or meaning, quite the reverse. What i was trying to say is you have to be careful how you present the data to the NN. For instance if you use shots on goal as one of your criteria the NN does not know what shots on goal are and purely treats this as a bunch of numbers. For starters how do you represent the match result. Do you use goals for and against and find a way of telling it who was at home cos it has no means of knowing this. Alternatively do you use a points system 0 for a loss, 1 for a draw and three for a win. If so how will the NN view this info, will it think a win is of three times greater importance than a draw. This is the sort of thing i meant. To be honest if i didn't get logical output i would have given up on the NN route long before now.

[oddsfellow]

Re: Neural Networks as a means of prediction. Merlin, to be honest i dont see abrasive comments just a valid link followed by the summary of results. If they are bad surely its better to say so?

[merlin]

Re: Neural Networks as a means of prediction. Oddsfellow, it's a carry over from another thread.;)

[dietcoke]

Re: Neural Networks as a means of prediction.

Hi Muppet just what can I say.You seem intent on presenting an abrasive mannerism both on Mr.Onemore's thread and now on this one.:hope

Merlin, I think you've got the wrong end of the stick there, I can't se anything abraisive in Muppets post. In fact the site looks pretty good with some good links to sites on AI and neural network research. I've viewed it quite often in the past and the results are not consistently brilliant for any leagues.

[snyena]

Re: Neural Networks as a means of prediction. I know nothing about Neural Network Systems, but here are my observations and thoughts about betting for what they're worth: 1. There is sometimes great value to be found after an upset happens (or for another reason). An example of this would be last year's odds of 1.5 on Latvia away at Liechtenstein after Liechtenstein had drawn with Portugal. It is odds like these that offer extreme value and while it is generally thought that yield above 20% is impossible in the long run, there are people and groups of people who pull this off and find 50 to 100 such bets per year. I'd be very interested in a Neural Network System that could spot such extreme-value bets (instead of pissing with many 2% yield bets, although that ain't bad either :) ). Is it possible? 2. I think a lot of punters read too much into 'form', various numerical analyses, etc. when there is a simple numerical stat, the team's league position, that fairly accurately reflects a sum of many small variables. I find focus on the price in combination with the league standings more important and more profitable than various partial analyses of i.e. head to head stats, last 5 games, and so on. The implication for a Neural Network System could be that less is sometimes more - simplicity might be the key. An example of this would be a system that scans various leagues (the more obscure the better :tongue2) and finds games where there's a huge table difference between the teams and the favourite is overpriced. So much for my thoughts... probably not much to do with Neural Networks, but perhaps helpful anyway.

[muppet77]

Re: Neural Networks as a means of prediction. sorry merlin. just wanted to show their findings for one of the most popular leagues in the world. reckon it's an informative site. loads of links to software and other AI sites. sorry mate didn't mean to be abrasive!

[Osesame]

Re: Neural Networks as a means of prediction. Why not use real intelligence, we all have brains.

[mr onemore]

Re: Neural Networks as a means of prediction. I had a small look into this - and tried some of the free downloads available on the net. I think it's quite an exciting development - but I really don't know enough to use them with confidence. My own feelings so far are: Totally agree with post above - garbage in, garbage out. Have to be very careful about the kind of data used. My biggest worry - is that even if you do have a NN performing in a positive fashion, the algorithm produced remains hidden - at least it is not something that can be analysed in an easy way. One of the tenets of science is that a formula/law/rule can shed further insight on the problem, leading to new or better understanding. With the NN, when it doesn't work, it's not at all easy to see why. And when it does work, still don't really know why!! Putting my own money on a bet in these circumstances is a non-starter. I had the impression that you just keep trying new combinations of variables until you find one that works - it's just a whole different approach. Still - interested in this kind of stuff - would like to see someone who's had a real go at this post up their actual methods/results.

[PAULM03]

Re: Neural Networks as a means of prediction. You're absolutely right mr.O, Neural Networks are a 'black-box'. I've contemplated using them for my own work but have never had the time to do it properly so it's fallen by the wayside. 'Fuzzy logic' and Genetic algorithms are 2 modern techniques that may also be useful to apply to betting, but again at the minute i have neither the time nor inclination to open a can of worms. Good luck:ok

[merlin]

Re: Neural Networks as a means of prediction. Thanks Muppet I have no desire to be at loggerheads with anyone, sometimes it is difficult to read into messages just what makes the other person tick.In the interests of finding true and promising systems I think that it is time to move on and leave our fairly 'stick' start and put it down to a misunderstanding.I hope that we can be of mutual assistance in exploring the avenues that no doubt need to be explored and discover further aids in the world of prediction.Have you yourself had any experience of NN's, as I said before it is not all doom and gloom, and I have an open mind when using these.There are many variables one can use, and I am not denying that as a prediction tool in its own right, NN's are not yet worldbeaters in my book.But as long as there is room for scope, I will plod a furrow.Hope to be able to discuss this further with you and really exchange ideas.Cheers fella!:clap

[merlin]

Re: Neural Networks as a means of prediction. Hi there Dietcoke, thanks for your comments.We are now treading old ground as Muppet and myself have agreed to forgive and forget.I did get the feeling of some offending remarks, but this was on another thread completely and nothing to do with his postings here. Anyway we appear to be playing Happy Families once again and I hope that there is no harm done and no ego's tarnished.I am aware of the many sites that have links to NN activity, but what I really wanted to do on this thread, was to really discuss openly punters personal experiences using NN's, the variables they used,and the predictions at output.If we can as a group have a fairly comprehensive consensus of just what variables and how many we should include, and the importance of those variables as a prediction tool, then I don't think that we will be too far away.So have you had any experience in using these,and what kind of variables would you view as being a candidate for inclusion, should you be using an NN?;)

[merlin]

Re: Neural Networks as a means of prediction. Hi there, and thank you for you constructive comments snyena. Dealing with your first point raised on the subject of value bets, preferably big value bets, and whether Neural Networks could possibly explore this avenue. Strange that you should say this, because as I speak a friend and I tried a new approach a couple of weeks ago.The Raw Data was fed into the system, eight variables in all, and initially we were looking for a fairly high percentage of correct predictions.The query file was prepared and the results were churned out.In terms of correct predictions, we found nothing to clap and wave our hands in the air about.However after close scrutiny, we found around ten of the highest priced homes and forecast as homes, turned out to be good priced aways.The same happened when we examined the highest priced aways and forecast as aways, they turned out to be home wins also at decent values.So it is still early days on this one,but we need to do some more exploring. As for point two Form and League Tables and whether we tend to place too much emphasis on form.Personally I think that form is more useful to us in the short term, say the last five or six matches.Otherwise it becomes too distant and of little guidance to us as a predictive tool.Form is a guide, and if we ignore, just what do we use as a substitute.As for League Tables, I am not a great lover of these as they are not a true reflection due to the different sides that have played each other at any given point.Probably they are best measured at mid point and at the end of the season to be of any use.So there is reasonable arguement to endeavour to include the average opponents strength into any ratings, or perhaps even home and away trends using the Rcaing & Football Outlook.Whether there is scope for a NN to scan all leagues is beyond me, and possibly could be done by someone with programming skills. From an ordinary NN point of view, I would say keep it simple.Cheers.;)

[przeszczepan]

Re: Neural Networks as a means of prediction. Hi Merlin I can see you're doing anything you can to get the thread back on the track of concrets. Still I hope you won't mind a few more general comments before you get the discussion on variables going. I agree to a considerable extent with what Mr Onemore has written.

One of the tenets of science is that a formula/law/rule can shed further insight on the problem, leading to new or better understanding. With the NN, when it doesn't work, it's not at all easy to see why. And when it does work, still don't really know why!! Putting my own money on a bet in these circumstances is a non-starter.

Weights one receives having taught a network are no interpretable. Yet worse- it's not possible as far as I know to examine their statistical significance. In fact one is not able to run any test the way it is possible for a traditional econometrical approach thus it is not possible to determine what is going wrong with the current network and what steps might be taken in order to eliminate the errors. I've written not much is left for ANNs from my own experience- results were at most comparable to those of an econometric model- if they were better, it soon turned out to be the case of overfitting. I am not saying it cannot be done- my use of the ANNs was constrained to a somewhat blind-like trials of a software's capabilities- there were so many types of networks available to choose from + so many learning algorythms I didn't really have the scientifical background to decide which one to take advantage of. I suppose it would be of much bigger benefit for someone more proficient on the subject.

[muppet77]

Re: Neural Networks as a means of prediction. hi all. good ideas so far. i'm not that interested in using NNs, just the thinking that goes into them. I know next to nothing, but if anyone wants to drop me a pmail, i'll give you the email address of a friend who knows a bit about them and has had some experience.

[merlin]

Re: Neural Networks as a means of prediction. Thanks for your comments Osesame and your question on why not use real intelligence. Nothing wrong in using the human brain it is a truly wonderous piece of our everyday functioning, and without it, we would be lost.However when we refer to the likes of Neural Networks, the amount of data handling and calculations that has to be dealt with is astronomical.The raw data we input all has to be dealt with in a methodical manner and then becomes a learned situation each time one match result and its accompanying data is produced.In most NN's there are generations upon generations learning from that which has gone before,producing and counter producing over and over again. This I feel would be too great a task for the human brain.;)

[przeszczepan]

Re: Neural Networks as a means of prediction. In fact, as far as I know, human brain is quite a decent piece of a neural network in comparison to what peple are using as artificial substitutes;)

[brookes]

Re: Neural Networks as a means of prediction. A very interesting thread with ideas that could result in a long term profit. I also agree with Osesame.

Why not use real intelligence' date=' we all have brains.[/quote'] My biggest problem with my brain is that it has a bias towards certain teams , sometimes with no logic to it .Long ago I have learnt not to place bets on my team. So now the bias go to certain other teams, which can result in a good run of profit then a long run of losses all due to the decisions made by my "black box[brain]" Cheers ,keep the discussion going

[merlin]

Re: Neural Networks as a means of prediction. A big and warm hello there Mr.O, indeed an honour to get you calling in on this thread, nice to air one's views.I totally agree with most things you have commented on, and can well understand your concern as a mathematician, that you yourself have little control over the algorithm, as this is often built into the system package.We as NN users and I think that it is true to say, we are also NN experimenters to a certain degree.Completely true in what you say that if one NN does not produce the desired results, try another.In my experience there appear to be many sets of variables that have produced different results over the years, but there also appears to be no answer,other than trying something different and hope for better things.From your point of view Mr.O probably not very satisfying,but I suppose it becomes part of the overall learning curve at the end of the day.It is mooted that we learn something every day, so I suppose it can all be classed as part of the enjoyment.Two weeks ago we built Ladbrokes odds into the system and this had not been attempted previously, with encouraging results.The art of discovery.I can also understand you cautious approach to putting your hard earned cash into NN predictions, at this point in time I would not do this either, it would be as it stands, a quick way to the poor house!Hope to continue this further.;)

[merlin]

Re: Neural Networks as a means of prediction. Hi there PaulM03, nice to see your comments also.Keyboard is absolutely red hot in trying to answer everyone in turn, a bit overwhelming.Anyway good to see that you have an open mind also in the uses of Neural Networks as a means of prediction.I continually search along with no doubt countless others for a method which will I hope produce a decent percentage correct of all matches predicted.I currently have three different models, which I dabble with from time to time continually looking for improvements. Your ideas on Fuzzy Logic and Genetic algorithms are on the right track.One of my software packages boasts twelve different NN models including the two quoted above, but have still to experiment with the majority of these.Looks like you are strapped for time in using models of this nature, unfortunate really,because it does become absorbing but very time consuming.Still, good to have your comments,please call in again.Cheers.;)

[przeszczepan]

Re: Neural Networks as a means of prediction.

we are also NN experimenters to a certain degree.Completely true in what you say that if one NN does not produce the desired results, try another.In my experience there appear to be many sets of variables that have produced different results over the years, but there also appears to be no answer,other than trying something different and hope for better things.

Btw, what seems to be something like a new trend in econometrics is grouping all variables you might possibly think of (spots on the Sun as well as frequency of wisiting the toilet by your neighbours' daughter qualify) and building a model on that. May seem crazy but people who rather know what they are doing (and do have valid scientifical arguments for that) use it extensively so there might be something in it. Might be worth a try to employ the strategy with an ANN as I guess it's better at weeding out useless inputs. Just a thought.