Neural Networks as a means of prediction.

[oddsfellow]

Re: Neural Networks as a means of prediction. The "real intelligence" theme is an interesting one. Automation is often used in everyday life to reduce errors and also reduce the amount of work done by humans and i see NN's in the same manner. I am a sucker for using examples so here goes. Take any match you like and think about the factors involved before you make a bet. Recent form, long term form, goals for, goals against, opposition faced etc. Having obtained all these factual points individuals may create ratings, tables to illustrate their point then based upon the result of this exercise place a bet. The more realistic option is that the vast majority digest this information in a visual manner only (tv, text etc) and go purely with gut feelings. For the more learned lets assume the first option is used. What weighting do you give to say recent form in the decison making process? How do you know this is correct? Some may say here you can check back historically and use this as a guide which is true to an extent but by and large it is guess work. However what if you had a program that once you put the various inputs into it, it would automate the task of deciding how important these were in the output (result) of the event and furthermore provide predictions for future events based on the same inputs. NN's are not everybodys cup of tea and sure aren't the easy things to get to grips with but i believe they do have benefits.

[merlin]

Re: Neural Networks as a means of prediction. Hello there Oddsfellow,apologies for missing out a reply to your post number 11, which I thought was constructive.Sorry to have mislead you on the question of data presentation.Obviously you have been dabbling with NN's as some of the basic questions are being raised, and similar in ways to those which I used to ask.As you said earlier the programme does not know in what format you are presenting your data,but I myself believe that where possible it is better to present this in a logical manner.Let's say that you had 9 variables four relating to the home side and four relating to the away side,with the ninth variable your match result.So obviously your output column would be column nine.My data would thus be arranged as the four home variables first,followed by the four away variables,followed by the output column.Makes sense and is easier for input purposes.So the data presentation at this point looks logical.The main question then,is what data do I use? Are shots on goal really measurable when it comes to the final result? And what degree of importance to they display in accordance with the other data?Goals for and against are really long and cumbersome, would it be better to use Average goals for and against and so keep the percentage data at a lower level?You have also asked how the NN determines the value of the result, answer is it doesn't.It retains all the lines of data thus entered in the input stage, stores and remembers these as a series of events yesterday,last week or even months ago,and when a similar set of variables and data is produced,it says something like, " Hey, I remember this,the result was....." Works something like that, so your 0, 1 and 3 could well be used.We need to establish by discussion, what we consider to be important in predicting the correct outcome at the end of the day.Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction. My thanks to you Brookes for chipping in with your thoughts on 'the little black box' theme.Perhaps the brain is as you say a Neural Network in its own right.Goodness knows how many calculations and judgements it has to perform in an average lifetime, so altogether an ingenious piece of performing equipment we have housed in that 'thing' perched on our shoulders.Neural Networks appear to be an extension of the brain by the way in which it deals with the countless numbers and masses of data presented to it on a daily basis.Maybe we are all a wee bit lazy in the way we utilize ours, hence the need for something a little different - the NN.I have found over the years nothing is handed to you on a plate, and many argue there is no true way to control the algorithms etc,but the only way forward in this field is to endeavour to be a bit of a pioneer, with further experiments I am sure improvements can be engineered, and if we did not have the light bulb, we would all be sitting here in the dark! Cheers.;)

[Osesame]

Re: Neural Networks as a means of prediction. I still don't see how neural networking can be to any advantage. There are so many variables.Can you decide which to include? The weather I believe plays a part, it may only be a small part,but how do you input weather into a NN. The time of day, the day of the week,injuries, suspensions,new players,new managers,financial crises,all seem to play a part.

[merlin]

Re: Neural Networks as a means of prediction. Hi there to you przeszczepan(nearly took me ten minutes to type your name lol) Many thanks for your posts #23,#26 and #30 and I have noted your reactions to the lack of control of your NN's.I have also noted your remarks on analysis relating to scientific backgroud etc and as I have already stated previously, there does not appear to be a constructive answer on how to go about getting the desired results using scientific methodology.The NN's usually have their own algorithims etc built in, so it looks as though it is up to the user/s to find their own way forward.I think that you hit the nail right on the head in mentioning the way we use our data, just what is the statistical significance we are trying to achieve by means of this network?I have read several instances here as saying something like,"What is wrong with my network?" Ask yourselves this question,"Am I feeding relevant material into the network to achieve the desired result?"We have not yet entered into discussions on this, and no doubt when we embark on this, there will be interesting thoughts and ideas on the subject.As for your spots on the sun correlating to visits to the toilet, the answer lies out there my friend.Thanks.;)

[spreadman]

Re: Neural Networks as a means of prediction. Dear All, it's great to see such a good response on this thread . It's just a thought , but to make things even more interesting, what say someone who can produce predictive data from using a neural network programme, posts their predictions, and other members,(non 'nns'), provide their predictions for the same events? I'm not sure if this would prove anything conclusively, but could provide an on-going thread branch of interest to all, 'nn-ers' or otherwise. Any views? Or am I talking out of my rectum here? :moon Spreadman.

[merlin]

Re: Neural Networks as a means of prediction. Great input Osesame, you are really getting into this! You have started to ask yourself the important question on just how do you classify the many variables that exist and could be included within any NN framework.The weather is a perfect example, is it fine, is it rainy or is it even snowing, and what if....... How does a 12 o clock kick off differ from a 3pm or 5 pm kick off, and just what effect has injuries and suspensions have on the NN output?Now if you go a little deeper, ask yourself are any of these measurable and if so, how do I go about measuring these with a view to input?The answer I think you will find is that they are not a significant measurable factor and of little use to NN prediction. We could go on forever with such as, "What does Ryan Giggs eat for his breakfast?" or did Joe Cole break the curfew and get to bed late.We have to determine just what is measurable,and what is considered to be useful data in the NN prediction process.;)

[Osesame]

Re: Neural Networks as a means of prediction. The answer is that I believe they are significant but not measurable.If you are unable to input data that is significant then surely the whole thing becomes pointless.

[merlin]

Re: Neural Networks as a means of prediction. Thanks Spreadman, and an interesting thought of yours.From my point of view I don't think that the time is right to start posting predictions as from what I see of previous posts, I get the feeling that NN predictions are not going to break any eggs with sticks, many problems seem to still be out there along with doubts on the capabilities of NN's in general.If anyone would like to take up the gauntlet, then by all means do, would be good to see and no doubt create interest, but at the risk of reapeating myself there looks to be more of a case currently in looking for value bets from my point of view, than from setting the world alight with predictions.I think that in the course of the next few posts, we will be looking for input on thoughts and ideas on what punters consider to be the essential ingredients, variables etc, which should be included in the decision making process.;)

[merlin]

Re: Neural Networks as a means of prediction.

The answer is that I believe they are significant but not measurable.If you are unable to input data that is significant then surely the whole thing becomes pointless.

originally posted by Osesame. I would go along with you on this Osesame but should we not then consider other measurable alternatives which may move us towards the desired objective? Packing in would seem to be a fruitless way out, besides one's curiosity and strive towards greater results would provide the driving force for me.;)

[Osesame]

Re: Neural Networks as a means of prediction. There are just too many variables in footy betting.I just don't see how a NN can "understand" that one top striker is injured ,one is on suspension and the reserve goalie is playing. To me these are significant facts that the NN between my ears can interpret.

[Osesame]

Re: Neural Networks as a means of prediction. I am not trying to dissuade you from from trying, but really saying that the current factors have to be taken into account after a prediction is made.

[oddsfellow]

Re: Neural Networks as a means of prediction. Osesame, I see exactly what you are trying to say and it is a concept i struggle with personally despite using NN!! The way i had it explained to me where if the inputs for the NN provide Outputs that are reliable then the other factors no matter what you think do not have such a bearing on the result as one might imagine. I still struggle with that but i tend to keep an open mind. I use it not as the definitive answer but as a way of highlighting initial value from which i then analyse the matches concerned for the various intangible factors. Also with a bit of imagination you can include virtually all factors you want. For instance first team players missing through injury or suspension can be represented by a numeric value (doesn't tell you there star quality admittingly but it can be measured in this manner), is it windy or not can be represented by 1 or 0, Managers time in the job could be expressed in days. However having said all this i personally limit the factors to a few and wouldn't recommend overloading it as is comes back to GIGO else.

[Osesame]

Re: Neural Networks as a means of prediction. My belief is in studying market forces to make predictions, surely the combined efforts of all the bookies and other knowledgeable people can find a few winners over a weekend.

[oddsfellow]

Re: Neural Networks as a means of prediction. I wouldn't disagree with that approach.

[ONEDUNME]

Re: Neural Networks as a means of prediction. Hello guys - I must confess that I haven't got a clue what you're on about but there seems to be a lot of words in this thread and very little substance so far. How about putting it to the test with some real predictions. Then we can cut the airy fairy debate and get down to real life?

[spreadman]

Re: Neural Networks as a means of prediction. Well spoken Onedunme. I suggested this in my previous post. Might I be so bold as to put myself forward, provisionally, to represent the non 'nns' - not that I don't believe in the potential of nn's, and their predictive qualities, but in order to earn credibility with all punters, and hopefully be of use to us all in relieving the bookie of his cash, some quantitative proof for all to see would, I believe, resolve this matter, and be of benefit to all. Please don't get me wrong - I am not 'down' on nns, but by adding 'substance' - 'fleshing out' the bare bones - the 'non- believers' will become, hopefully the converted. If I get positive feedback, maybe the moderators would consider some input, e.g. setting parameters for types of predictions, reasons for choosing them etc. Spreadman :ok

[przeszczepan]

Re: Neural Networks as a means of prediction. Merlin, I don't remember participating in such a well managed thread- well done:)

I have also noted your remarks on analysis relating to scientific backgroud etc and as I have already stated previously' date=' there does not appear to be a constructive answer on how to go about getting the desired results using scientific methodology.The NN's usually have their own algorithims etc built in, so it looks as though it is up to the user/s to find their own way forward.[/quote'] I believe there are some guidelines one should follow pursuing a particular project. In other words it may be better to use a given type of network for a given type of data, choose a particular learning algorythm for a given problem, select a particular number of layers for a given number of variables and so on. I had no or very little idea about that to be honest so had to experiment a lot. Verification of a model is considered to be one of the vital phases before it's introduced as a means of prediction. It's were you run a bunch of statistical tests verifing any hypothesis you might have as for the model. One of the most important questions you should ask yourself is whether your choice of variables was suitable ie do the spots on the sun really influence your variable. You get the answer pretty easily having run a test on itssignificance. I know of no way it can be done for an ANN.

[merlin]

Re: Neural Networks as a means of prediction. Just arrived back in and tuned in to see the latest gossip.Looks like some comments have been passed between Osesame and Oddsfellow and some of their own thoughts on the subject.Osesame does not understand how NN's can identify injuries, substitutions and new managers, but Oddsfellow kindly points him in the right direction by advising that numbers such as 0, 1, 2, etc can indeed represent all kinds of situations,again it is down to how relevant they are for the model in question.As for making predictions Osesame I note that you study the market forces, which is not a bad thing.No one here is putting a case forward that the NN's are ready to take over,far from it.I think for those that are experimenting with these will appreciate some good hearted banter, seek some of the group's ideas and perhaps even variables when we get around to discussing these.The purpose here is group discussion with perhaps benefit to those users at the end of the day,in being able to get better use out of their particular system.But good to see constructive input.;)

[merlin]

Re: Neural Networks as a means of prediction. Hi there Onedunme welcome aboard.Sorry if you see very little substance so far, but the purpose of this thread when first raised was to invite those interested in Neural Networks, and to hopefully attract a few who know nothing of these, into some discussion, with a view to attempting to better our models and perhaps learn from one another.There was no mention of hands waving in the air as a campaign on behalf of Neural Networks, let alone a "Gunfight at OK Coral" to decide the outcome of NN's v Punters predictions.There seems to be a few around who want the guns drawn and the bullets flying!:D I believe this thread has to walk before it runs, judging by the contributions so far, and good contributions at that,I think that we are progressing in the right direction.Maybe the shoot out may come later when we have exhausted all avenues.;)

[PAULM03]

Re: Neural Networks as a means of prediction. Does anyone have any NN software here? There's a free Matlab toolbox here: Netlab But you obviously need matlab first (which i have). Haven't downloaded it yet and not too sure how user friendly it will be. Does anyone have any links to anything similar?

[merlin]

Re: Neural Networks as a means of prediction. Hi there Spreadman, and thanks for volunteering your services as a non NN predictor.However, unless there is an absolute NN wizard out there among you who has an amazing predictive model at the ready and willing to take up your challenge, then you will be firing those bullets into thin air.They say patience is a virtue and how true that is,so I suggest we hold fire on that one, no one is going to run away and certainly nothing spoiling.When this thread was first started there were no plans announced for an 'us' and 'them' scenario, all that was envisaged was that those NN experimenters ploughing a lone furrow, could unite in common purpose and perhaps learn from discussion and the varying viewpoints.To those NN experimenters I say welcome aboard.To those who know little but are willing to keep an open mind, again I say welcome aboard.To those who know nothing and have no clue to what is transpiring, I say re read the threads, you only get out what you are willing to put in, welcome aboard.To all others with constructive thoughts and ideas, again I say, welcome aboard.:clap

[merlin]

Re: Neural Networks as a means of prediction. Hello there przeszczepan and many thanks for your kind remarks, I am glad that we have someone out there who is enjoying the thread and its comments.As long as we all remain constructive and present our viewpoints in a logical manner I am sure that it will be of immense benefit to all who participate.It's early days yet,blimey we have hardly taken off yet,there is lots to discuss yet I feel.What NN package are you currently working on?How many variables are you using on the model, and have you had any success? Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction. Hi there Paul I am pleased to see such an early return.I see that you mention Matlab, never heard of this,what is it,what does it do? In answer to your question on NN software, I currently have three:- EasyNN, Neuroforcaster,which was quite expensive, and Decision Maker. I experiment with all from time to time, but have not drawn too many firm conclusions yet as to which is the best model. EasyNN is relatively cheap and support is good for those interested.I think from memory it is around £19 or so and can be downloaded for a trial period at www.easynn.com Not my site by the way, no plug here,just for interest. Thanks for your visit, and hope that you have the time to investigate further. Cheers.;)

[PAULM03]

Re: Neural Networks as a means of prediction. you've never heard of matlab:eek Probably the most popular piece of maths software in the world. have a look here matlab

[merlin]

Re: Neural Networks as a means of prediction. Thanks for the grand tour PaulM03 means little to me I'm afraid.Looks like more for up and coming programmers.Never had the aptitude nor the interest in this sort of thing, I tend to be more into rating/prediction systems.I would need to be enlightened further on the likes of Matlab.Cheers.

[przeszczepan]

Re: Neural Networks as a means of prediction.

What NN package are you currently working on?How many variables are you using on the model' date=' and have you had any success? Cheers.;)[/quote']I was experimenting with Neurosolutions for Excel (a bloody powerful package and as expensive I guess) at school. There is a free trial version on their site. As for the number of variables it varied from 2 to about 14 as far as I remember. Success? Hmm... just decent predictions, however not good enough to beat the line so I guess the answer is 'no'.

[merlin]

Re: Neural Networks as a means of prediction. Hi there przeszczepan(going to call you pp from now on)I note that you have used Neurosolutions for Excel while at school.Had a look at their site you so kindly referred to,and the product looks pretty good, but also pretty expensive, seems to be $250 for that model, which puts it into the bracket that my most expensive NN cost, which I think was somewhere around £295 when I first got this from the Far East some five years ago.That is Excel based also, shows results nice and clear, but must be saving files wrong, as I tend to get error messages from time to time.Currently I use around nine variables with the output being at column nine.I think that the time is fast approaching when I think that we may start considering data variables and how we think that they would interact with the various models.Keep thinking...... keep talking!Cheers.;)

[spreadman]

Re: Neural Networks as a means of prediction. Dear Merlin, didn't mean to create an 'us' and 'them' scenario, I just felt that more members would be interested, nay, entertained by a 'gunfight' (with blanks, I hasten to add!), on your evolving thread. I agree it is still in its infancy, and no one has surfaced yet with concrete variables, predictions etc. but it could be a direction we may find ourselves, voluntarily or otherwise,drawn towards. I, personally, am not on a witch hunt for or against nns, but my ignorance at the moment only allows me to form a 'fuzzy impression' of nns and their uses. My fault I know, and I will endeavour to educate myself more thoroughly on this subject. Are there any websites suitable for a dimlow like myself that explain nns in monosyllabic terms? With reference to 'input variables' if that's what we are waiting to discuss, then I can suggest one or two that I consider relevant to spread betting, if you so wish. As I have stated previously on this thread I am certain that the spread companies use nns to assess and evaluate their spreads, before releasing them to the market. Keep the input flowing, there's no live Sky TV football until Friday so I have time on my hands. Spreadman.:loon

[PAULM03]

Re: Neural Networks as a means of prediction. mmmm, there's a neurosolutions for matlab too but its $95. Think i might give the netlab one a try seeing as it's free. Looking forward to discussing some variables:ok