Neural Networks as a means of prediction.

[merlin]

Re: Neural Networks as a means of prediction. Looks like we have you licking your lips PaulM03.Moderator or no moderator, looks like you are getting interested..... and that for me ..... is great! It means that we are starting to achieve what we set out to do,.... to get punters to open their minds on NN's and talk to one another.Really wonderful Paul.Great to see you finding the time to explore further..... I'm all ears!!:clap

[muppet77]

Re: Neural Networks as a means of prediction. wow. just downloaded the free trial version. looks impressive. can anyone explain how to run a basic test , without me spending ages figuring all the demos and menus out, so i get an idea of what it can do interms of footy without looking at the gender of crabs tutorial - just want to see what it can do????

[merlin]

Re: Neural Networks as a means of prediction. Hi there Muppet young fella, good to hear from you again.What exactly have you downloaded Netlab or EasyNN? Usually you have to trundle through the help section to get anywhere,so will await your further comments on this.;)

[muppet77]

Re: Neural Networks as a means of prediction. Neurosolutions. gathered that i have to save data as text files, but i tried running it, but just got error messages. have you used this one?

[merlin]

Re: Neural Networks as a means of prediction. Had a quick download of your same version Muppet, but I don't know whether it is the lateness of the hour and I am getting tired, but looked pretty complicated as I browsed through some of the demos.Gave up in the end and removed the whole download.Usually the files saved do make reference to txt. files.In one of the posts above przeszczepan mentions he used this at school, so perhaps he can help on this one.Cheers.

[merlin]

Re: Neural Networks as a means of prediction. Looks like the sun has gone down on the OK Coral, no guns blazing, and all quiet.Looks like a good time to examine what sort of data we could have at our disposal.So I will throw the following variables out for your further consideration gents.Whilst examining and studying these, try to be subjective and think in terms of how useful these variables should be within the prediction process.It is obvious that we should keep the number of these down to a manageable size to enable the network to deal efficiently with the data being presented.It should be meaningful and logical and I would suggest that we try and keep the home data to the left and the away data to the right, with probably the final output column on the extreme right.If there is anyone who does not follow this logical process or has other ideas on how this data should be presented, then by all means let's hear about it!So before I present my list of variables, there will be some there that I have put in as Red Herrings in my humble opinion, just to stimulate conversation, I myself would never dream of some of these for inclusion.However, again if you feel that there are others of greater importance, and not on this list, then again,feel free to expand.So OK it's blast off time, here is my list for your further study and comments:- Games Played H/A - Games Won H/A - Games Lost H/A - Games Drawn H/A League Position - League/Division - Average Goals Scored - Avg Gls against Avg opponents strength - Attacking Play - Total Goals For - Total gls against Shots on goal - No.of corners - Playing conditions - Playing surface - Injuries/Suspensions - Manager - Kick off time of day - Referee - Penalties - There is a fair size list there to consider and in turn I hope they will attract some further discussion and more importantly to get you thinking in what best fits the NN output requirement.Finally before closing I have left this until last. Can we further make use of other types of ratings? Take the Racing and Football Outlook Index Ratings as a prime example.They have excellent ratings compiled on a weekly basis which could prove useful to our cause.They have Index Ratings for both home and away matches for all teams up to and including Nationwide Conference.Also there is a current Index rating produced.Further to that we have home and away weekly trend figures to help guide us as to just how each team is currently playing.We also could make use of Bill Hunters Yearbook ratings for those who have his excellent yearbook ratings.(Anyone who wants to know more give me a shout by all means)(To moderators,this again is not a plug,just a further set of stats which could help the cause)His ratings include team points differences, and again it is food for thought in our furtherance of NN rating material. So I throw this open to the floor and hope that we can stimulate your further thought processes.Cheers Gents.;)

[merlin]

Re: Neural Networks as a means of prediction. Sorry, I forgot to include one other variable for consideration :- Is there a case for using decimal Bookies odds into the NN?;)

[oddsfellow]

Re: Neural Networks as a means of prediction. Merlin, Credit to you for keeping the thread going :clap This is only meant construcively but i think the use of bookies odds and ratings is totally the wrong way to go. NN's work best when you give a list of inputs (factors) that produce an (ouput) result. Furthermore if possible each factor in order to keep the model simple should not overlap another. For example league position and games won is pretty much a measure of the same thing. A team with a high league position will also have won lots of games! I think its important to remember as i have stressed before the NN does not know about any inter realtionship between factors and it is important to have clearly defined inputs. Which brings me onto ratings and betting odds. Betting odds are by far the most accurate predictor of football matches, fact. However they do not affect the outcome of the match. Bookies odds will be based upon league position, form, attacking strength, injuries etc all the things you are using as inputs. By including them in any NN it will see that the odds are of significant importance to a result of a match which has no control, and your NN output will by and large mirror the bookie odds. Its pretty similar with ratings.

[mr onemore]

Re: Neural Networks as a means of prediction. Well, my humble advice here is that I would say you should never factor the bookies odds into your ratings - if that is what you want to go up against later...They should quite literally be the last thing that you compare your ratings/predictions with if you are looking for a system to beat the bookies odds. As to the other list of possible indicators, I'm still trying to get to grips with the concept of NN's and how they work. I still haven't really come around to the idea that you try several different combinations of variables at a time and throw them into the mix - allowing the NN to decide which are more important than others. I think you have to be careful that your variables are truly independent also before the NN gives undue weight to a farily insignificant data item.

[mr onemore]

Re: Neural Networks as a means of prediction. Odds - you beat me to it whilst I was typing ;)

[PAULM03]

Re: Neural Networks as a means of prediction. Just an initial idea which may be rubbish. Why not have each previous game played as a separate variable? then let the NN apply the weightings as it sees fit. For instance, as we're now approaching the latter part of the season the first game back in august should have little bearing on a game played this weekend, but recent games may have some significance. Since the strength (as i see it) of NN's is their ability to recognise patterns and apply weightings to variables one may be able to apply a more realistic weighting to recent games and find a better balance between 'form' and 'class' than by a purely arbitrary or subjective method.

[oddsfellow]

Re: Neural Networks as a means of prediction. Complete list of commercially available software can be found here http://www.faqs.org/faqs/ai-faq/neural-nets/part6/section-1.html and free stuff here http://www.faqs.org/faqs/ai-faq/neural-nets/part5/section-2.html

[merlin]

Re: Neural Networks as a means of prediction.

This is only meant construcively but i think the use of bookies odds and ratings is totally the wrong way to go. NN's work best when you give a list of inputs (factors) that produce an (ouput) result.

originally input by Oddsfellow. Thanks for your input Oddsfellow.Probably when you say Bookies odds and ratings, you may have got the wrong idea.The Bookies odds is correct, but there is no mention of using bookies ratings.So why should we not go about using these odds as a yardstick?As decimal odds they are input factors and they do have measurability in the way that their chances are considered to be for that particular match.These are shown as chances of home win,draw and away, and when included in the NN they do participate in the end product, the output.These have been put to the test along with other factors, and gave encouraging results, not by way of correct predictions I may add, but by highlighting value bets which were originally the weakest home odds forecast as home wins, but ended up as away wins when the results came to hand.So not all doom and gloom there.I am not quite with you on the inputs being used in our NN as being the same thing that the bookies base their odds on, surely this is for each individual to follow his own beliefs in those variables which he thinks should be included in the model, and should differ in many ways to those being utilized by the bookies.;)

[dietcoke]

Re: Neural Networks as a means of prediction. Possibly, along with our discussion of inputs we be should be considering where and for what leagues the suggested information is available. For example, I think that key players availability is a metric of major importance but teamsheet informtion is not easy to come by except for the premiership and you have to able to collect it, do your analysis and place your bets in the 40-60 mins before the games. Anyway if there are 50 games on saturday where can this info come from? if it cannot be obtained there is no point in using it as an input.

[merlin]

Re: Neural Networks as a means of prediction.

Well, my humble advice here is that I would say you should never factor the bookies odds into your ratings - if that is what you want to go up against later...They should quite literally be the last thing that you compare your ratings/predictions with if you are looking for a system to beat the bookies odds.

originally posted by Mr.Onemore. Thanks again for your visit here Mr.O, and thanks for the very welcome input. Your opening gambit re not factoring odds into your ratings, I take it you mean the NN.I don't think that I myself will be looking to compare bookies odds by this method, but please refer to the comments in my reply to Oddsfellow on that section.As far as the NN goes it is a continual learning process particularly those NN's where they use generations of learning to build on.Not sure on your comments about throwing several different combinations of variables into the mix, do you mean all at once, or at vaying attempts?The more you tinker with these,the more you learn, and the more determined you become to find the answer.Cheers.;)

[spreadman]

Re: Neural Networks as a means of prediction. Thought I would put my penny'sworth in early at the start of what will prove to be a busy trading weekend. I have been thinking about my own personal 'variable inputs' - things, factors, I tend to think about, systematically, (subconsciously?), that I use each time before I place a spread trade or bet. Yes, I will use most of the variables previously mentioned,( spotted your red herrings, Merlin!), - ( you forgot one - whether D. Beckham is wearing a thong or jockeys), but slightly more contentious, nay, contraversial, (?), are input variables such as 'shots on goal', ( in my particular case, for total goals trades and total corners, - in fact, any goal related markets, I suppose ); 'history' : 'number of corners/bookings/ shirt numbers' in recent matches for these teams; 'result of previous fixture between the teams', ( probably more useful for spreadbettors in that if bookings were high last time, this could indicate the prospect of a 'dirty match'); referee's position in the 'strictness chart', (bookings again); team's position in 'fair play' league,(ditto), what I call 'last minute panic' data : change of player's shirt number, 'Hotshot' on subs bench; and, finally, though I'll probably think of more once I have posted, 'seat of the pants data' for in-running trades/bets: e.g. substitutions, injuries, players throwing rattle out of pram, personal vendettas developing 'in-play', and the 'spit-at your- opponent', and/or 'nut him' input factor,( you know who you are, Mr.Diooooof ). Well, excuse my rambling - just a tiny insight into the weird and wonderful world of spreadbetting on a foggy bank holiday day in not so sunny Jersey. I would be interested to hear punter's comments, observations, criticisms, abusives, death threats etc. Have a good Bank holiday, luck in your betting,(try factoring that input!), back later on my Class 'A' Mercedes thread with my bet predictions to try and lose my Easter egg money, ( apologies, Merlin -it looks like I just plugged my thread on your thread - is this illegal behaviour, Mr. Moderator, sir?) Yours, apologetically, Spreadman,( for it is he.....) :spank

[mr onemore]

Re: Neural Networks as a means of prediction. Hi Merlin, Regarding using the bookies odds as an input variable in your NN. As I see it - you will want to use your NN to make predictions about a forthcoming match. It will be important that you are able to formulate predicted odds for this match. It is only at this point that you will want to compare your predicted odds with the bookies odds - to see if you have a value bet. It's the same argument whenever you make a bet - formulate your own odds before looking at the bookies odds - to see where the value is. If you look at the bookies odds first - you can't help biasing your opinion. It will have the same effect on the NN. The other point was to make sure that the input variables you use are not dependent on each other, or just measuring the same statistic in a different way. As oddsfellow wrote - games won, and points, are pretty much the same thing - but the NN will give both a seperate weighting, distorting the process. I would suggest that your list of input variables be a diverse set of stats, but of course of a significant type. My own aversion to NN at present, is the use of several input variables when I'm used to using just one. But I'm certainly interested to see how the ideas follow here - as NN certainly work in some areas, so no doubt can provide a useful contribution here. By the way, the NN I experimented with last year was also the easyNN mentioned in a thread earlier - which has a free trial download. As a suggestion - can I ask that we use this for the thread so everyone who wants can follow the discussion. Or if there is a better free one - then let's use that. I'd be quite keen on experimenting with this idea.

[Dino]

Re: Neural Networks as a means of prediction. Good thread,just read through most of it. Agree with MO & Odds on two very important points. Using bookies odds in the input process is a very risky course of action.They are highly correlated to the actual match outcomes & therefore will be very influencial in the ouputs.Presumably the intention is to spot the small number of matches where the bookies odds are incorrect & this will be virtually impossible if you use those self same odds at the outset. Secondly using a number of similar parameters that are correlated with each other(shots on target & goals scored,for example).Again a very dodgy method.These correlations will never be perfect,sometimes they'll reinforce each other,sometimes they won't.Chosing to go with the best or satisfying yourself that the ones you uses are independent is vital,imo. Using outside ratings is better,but also problematic.You don't know any of the parameters used to generate the ratings for a start.I've never found the RFO ratings good enough to generate profitable value bets & the home/away split ratings are worse than the combined ones. Don't want to sound negative because there's alot of sense being posted & throwing away flawed ideas is as important as running with promising ones. Must dash.

[merlin]

Re: Neural Networks as a means of prediction. Hi there Mr.Moderator, alias PaulM03, thanks for your very welcome input. You mention trying to keep each game as one separate variable and then have the NN apply its weightings.I really don't see this as being practical and possible because as far as I am aware,the NN works from the data as input.Whether this is as individual input or clusters of input, would have to be determined, I really don't know.In relation to your comments on data input say last August as opposed to current data,surely what you are stating here,is that you prefer current form as opposed to much older data for your NN input.Usually I find that most NN's like a lot of data lines,and seem to work much better when this has taken place.It very much depends on what variables you have built into the network in order to discard the old info, as if you remember the network is learning from square one,day one and builds it structure from then.A complete learning process over and over again, remembering what has gone before.Besides you can always delete old data,cut and paste new data into the user file,and start off a new network at any given time.Good to see you thinking about it!;)

[Osesame]

Re: Neural Networks as a means of prediction. There is another variable that has not been mentioned, the "silly seasons".

[oddsfellow]

Re: Neural Networks as a means of prediction. Merlin, I'm not going to keep pursuing this line and will drop it after this post but just feel because it is a fundamental element of NN i would try and clear it up. Your inputs should be things that affect the result of the match, in what way do bookie odds influence the result? Answer they don't. As a means of prediciting the result they are excellent and thats what you are up against. How do you think the bookies formulate these odds? I would hazard a guess that the same data we use for our inputs is probably used to some degree in their compilation. Therefore if you want to include the odds as part of the NN why bother also including factors that have already been accounted for. Another point is, and no offence but i dont think a lot of people appreciate this despite me raising it several times, the NN has no idea in what context you are presenting the data. For example 10, 12, 2, 1.76, 2.25 with an ouput of 1 means what? No idea well thats exactly how the NN starts it doesn't know that you mean ave shots on target, off target, league position, ave goals per game, bookies match odds and match result as the output. As such it will see that the most reliable input is the match odds and base its output primarily on this factor. You might think well thats ok i will just be creating super efficient match odds but long term you won't. There may be statistical anomalies which may look like you will but long term you wont beat the bookie this way. Imagine for a moment in your dataset that you have 15 matches where the home team are 1.80, and by some coincidence they all win. The NN will view this data and realise straight away that if the home team are 1.80 the output is correct, so the next match that comes along with match odds of 1.80 it will overestimate its chance of success because this is what it has been taught. It has no idea that these odds basically represent % chance of success (if it did it would probably find it slightly confusing that using one of its inputs you then expect it to produce the same output ie % chance of success!!) I suppose what i am trying to say is if you want to create a system based around the use of bookie odds there are easier ways of doing this than using NN's. Furthermore you can say that the bookies odds are effectively a set of ratings. So by the same token as i think its fundamentally wrong to include odds i would also couple ratings in with this. Others may be of different opinions and i will of course respect that but my viewpoint has been established from significant literature and recognised users of NN's that i would certainly stand by it.

[PAULM03]

Re: Neural Networks as a means of prediction. excellent point well made, odds:clap This is how i got to thinking about taking each previous match as a variable. 1) The variables in a NN must be independant of one another (essentialy the fact just stated by odds) 2) Assume that the outcome of each game played is not dependant on the outcome of previous games This then provides a set of independant variables that may be able to be used to train an NN. How? i have absoluteley no idea. I've never used one and have only a very rudimentary understanding of how they work. Regardless of the variables used, might it be a good idea to just use a simple binary input for each variable? e.g. shots on goal. Does a team have more than 50% of it's shots on goal? the input would be 1 if yes, 0 if no. Of course the 50% is a purely arbitrary figure. In the same way the output could just be binary. 1 for a home win, 0 for a home loss or draw. The bookie's odds may be used as an estimate of the probability of a particular outcome, could they be used as some kind of threshold value in the NN, i.e. not as a variable. The way i then imagine this working is an NN trained using a set of historical data to determine if the probability of a home win is greater than the odds suggested by a particular bookie. How would i do this? fcuk knows.:unsure

[thanatos]

Re: Neural Networks as a means of prediction. I must be mistaken, but I thought a NN was not only made up of numbers by necessity? Would it not have to include names and 'features' such as 'shots on target' which are explained to the NN by means of interrelated numbers?

[oddsfellow]

Re: Neural Networks as a means of prediction. Paulm03, I dont think you are far off in your line of thinking. As i said in my first post in this thread i'm by no means an NN guru but have and do use them. From my experience the more black and white (1 or 0) you can make the data the easier the NN can interpret it. Some times this isn't always feasible but as a starting point its a good one. Your idea of independent variables and then letting the NN work out the correlation between the values is again sound. I promised i wouldn't mention it but if you are telling the NN via ratings that one team is stronger than another it will get the message, i prefer to let it work it out for itself as thats what you are asking it to do. The same goes for weighting certain factors, its very tempting but the NN will work out if the varaibles are indeed influential and adjust accordingly. Have never used EasyNN but will download that today and perhaps if we all did the same example with a set of data people may start to see the whole thing a bit more.

[merlin]

Re: Neural Networks as a means of prediction. Hi the Oddsfellow, been trying methodically to answer points raised and issues to hand in all posts todate.Will get there eventually I hope.Thanks for your list of Commercial and freebies websites in relation to NN activity, they may come in handy and useful to some of the punters here.Many thanks.;)

[merlin]

Re: Neural Networks as a means of prediction. Thanks for your input Dietcoke. Good point raised on how to identify leagues or divisions.These can be represented in a number of ways,calling Premiership 0, Championship 1,League Div 1 2 etc is one way.Using RFO index ratings can be another as the top teams are often rated greater than those below.As for player information, missing injured etc, I would think that it would be an enormous task to trace every team and perhaps every player in an endeavour to find out, especially if your NN time is at an absolute premium.I would tend to keep it as simple as possible for your needs.If your time is so limited then I think it would be difficult to do,and as you say if it can't be done in time for your bets,then little point in using this as a variable.Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction. Great post there Spreadman and some good variables well covered.I did forget to include David Beckham's 'undies' as a variable, pleased you brought this to my attention.:D I think that the gist of the matter is what one considers just what he wants out of NN prediction, then it becomes a personal choice as to those variables he sees which warrants their inclusion in the data.At this point in time,no one has grasped the metal and taken a lead.We have had individual members expressing odd items of data here and there,but there has not been one concrete case put forward as to how many variables they should include, and what those variables would be,and for what reason they should be included.For example,should we have ten,fifty or a hundred, or perhaps just three or four.Also should we ask ouselves, are we looking for a comparison of bookies odds in the final output,are we looking for correct match predictions,or indeed are we looking to try and predict the score?All food for thought!Cheers.;)

[PAULM03]

Re: Neural Networks as a means of prediction. Good idea about all downloading easyNN and using some collective brain-power to start to make inroads into how to actually implement a simple NN. Could we leave it till after the holiday weekend as i'm on an internet ban from the girlfriend till at least tuesday so my participation would be incredibly limited.

[merlin]

Re: Neural Networks as a means of prediction.

As I see it - you will want to use your NN to make predictions about a forthcoming match. It will be important that you are able to formulate predicted odds for this match. It is only at this point that you will want to compare your predicted odds with the bookies odds - to see if you have a value bet.

Posted by Mr.Onemore. Hi there Mr.O. welcome back.Regarding your quote above where you state that it is important to be able to formulate predicted odds for the matches in question.I assume you mean by NN prediction?I don't know whether that rings true with all NN experimenters, but as far as my predictions go, this is not the case in every instance.I keep trying varying methods,and more importantly always try to include those variables I think will make the difference from those previously forecast.As far as th actual odds go, I leave that mostly to looking at them outside of the NN framework to see just what can be spotted.Looks like you have spotted some of the Red Herrings I included in the list of variables.You are quite right, games won,points,position etc can be one of the same for our purpose,and highlights the importance of those to include.Good idea of yours about the EasyNN and whether anyone wants to refer to this.There is a trial version at www.EasyNN.com for anyone wishing to do that.Cheers.;)

[brookes]

Re: Neural Networks as a means of prediction. I am realy enjoying reading this thread even though I knew nothing about NN at the start ,now I've got some vague idea [is a little knowledge a dangerous thing ?]

.Also should we ask ouselves' date=' are we looking for a comparison of bookies odds in the final output,are we looking for correct match predictions,or indeed are we looking to try and predict the score?All food for thought!Cheers.;)[/quote'] The factors that would interest me would be the correct match predictions and if a system could predict the correct score all the better ! [what about goal scorer ?] The bookies odds compilers are experts at predicting the results are we trying to outwit them ? OR make money from them . Keep the thread going please ,it might be my lack of mental powers that makes me a "doubting Thomas " but I am willing to be converted.