Neural Networks as a means of prediction.

[merlin]

Re: Neural Networks as a means of prediction. Welcome aboard Dino and into the interesting discussions we are having re Neural Networks.I have noted also your comments on Bookies Odds and they fall very much into line with what a lot of the members have already mooted.I agree with your comments on the use of some variables as having imperfect correlations such as shots on goal goals scored etc, I never use these myself.I don't go along with your thoughts on not having confidence in outside ratings,unknown parameters etc.Is it vitally important that you should know how RFO make up their ratings for the purpose of NN data?Surely by the very fact that teams are rated in such a manner that if a team has a rating of 803 and its playing another with 699,then it is classed as being superior for that period.And how about the trend figures,are these also not worthy of inclusion,when they reveal how any team is currently playing?This could take the place of short term form within our network.Again,many things to think about.Thanks for the contribution.;)

[PAULM03]

Re: Neural Networks as a means of prediction.

.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.

OK, this is what i suggest we are looking for.

• Is the probability of a home win greater than that estimated by the bookies.

As no-one here appears to be an expert this keeps it nice and simple initially, a simple binary output, 1 for yes 0 for no. With regards to the variables, i think that these should also initially be binary, so we can start by asking some questions about a team: 1) do they have a positive goal difference 2) is their mean points per game greater than the mean for the division 3) is their mean points per game for the last 4 games greater than their average for the season 4) do they have more than 50% of shots on target All these questions are yes or no, and can therefore be described by a 1 or a 0. Feel free to add any you see fit.

[merlin]

Re: Neural Networks as a means of prediction.

.... in what way do bookie odds influence the result?

posted by Oddsfellow. Thanks again for your input Oddsfellow.Judging by what I am reading among these posts, I assume that a greater proportion of you appear to be looking to predict bookies odds from your NN data.Fine I say if that is what you are seeking.But as I said in an earlier post,not all of us are seeking to do just that.I think that we have to stop and think and try and read between the lines or by asking each other directly,"Just what are you aiming to get out of your NN?" You ask the question in quotes above OF, in what way do bookies odds influence the result? And I will answer in exactly the same manner as you did, they don't! I am not looking for any single variable to influence the result,but a combination of variables that interact with the network to produce a figure in the output column.How this figure is identified is by the criteria you yourself have laid down,groundrules etc.You kindly point out further down the post that the NN cannot identify variables as such, yet a paragraph or two later you say,quote: "The NN will view this data and realise straight away that if the home team are 1.8 the output is correct".How does it then know which one the home team is, and how does it know what 1.8 represents?Anyway young fella,indeed a good post with good points raised and indeed it stimulates discussion.Thanks.;)

[PAULM03]

Re: Neural Networks as a means of prediction. So, in simple terms what are you looking to achieve with your NN merlin?

[oddsfellow]

Re: Neural Networks as a means of prediction.

"The NN will view this data and realise straight away that if the home team are 1.8 the output is correct".How does it then know which one the home team is' date=' and how does it know what 1.8 represents?[/quote'] Merlin, the home team statement was a figure of speech so you are quite correct it won't know what it is, what it will know is that when this input (lets call it home team for argument sake ;) ) is =1.8 there is a 100% success rate. Um i wonder if it would think that this variable being equal to 1.8 is important :unsure You say you are not looking for bookies odds but if you are using it as a betting tool wether directly or indirectly you are by default. For example if you are looking at correct scores and it correctly predicts a 1-0 scoreline once every 6 times it is only useful to you if the bookie gives odds greater than that so by default you are looking at odds. Anyway i think as some have already pointed out this thread is getting a lot of talk and is probably in need of some action so i shall retreat and leave it to the very capable Merlin as i am beginning to echo.

[merlin]

Re: Neural Networks as a means of prediction.

This then provides a set of independant variables that may be able to be used to train an NN. How? I have absolutely no idea. I've never used one and have only a very rudimentary understanding of how they work.

posted by PaulM03. Another good post PaulM03 and I like the way you are thinking.Looks like we have no real experts on here,hence the reason for starting this thread,but I think that there is an awful lot coming out in the discussions.Again we have bookies odds mentioned in this post,but just what do you at the end of the day from your output column.Ask yourself the question,"What do I want my output column to represent?".Is it going to be match result,1,2 or X(3)?Is it going to be goal difference between the two sides?Is it going to be bookies odds? When you have an answer,then tailor your data to suit those requirements.You can download your EasyNN if and when, I think that this is down to each individual to experiment with data at their own leisure,could even get the girlfriend to type some data in for you!Cheers Paul.;)

[merlin]

Re: Neural Networks as a means of prediction. Welcome aboard Brookes,great to have a a new face.The thread was started as you will have probably gathered by now,to enter into discussions on NN prediction, thoughts and ideas, for those wishing to gain some knowledge,those with little knowledge,and those with no knowledge to hopefully be attracted enough to want to know more and start asking questions.I'm pleased that we have succeeded in getting you interested, and also pleased that you are enjoying the thread.As far as correct score prediction goes, I think it would be possible to maybe get near with total goals,never tried correct scores,and I think that it would be difficult to get goalscorers.Numbers would have to be generated for each team member for each team in question,for all those matches you are wishing to predict.Would take an eternity I feel.Let's try and creep before we walk.Keep it simple till you get the feel for these.Anyway great to have your input,keep it coming!;)

[merlin]

Re: Neural Networks as a means of prediction. Hi PaulM03 again another constructive post and some flesh around those old bones.So I note that you have put one bold foot forward and suggested what we are looking for:- Is the probability of a home win greater than esimated by the bookies? A good bold statement I feel, but why stop only at a home win? Why can't it be something like this: Is the probability of our match prediction,greater than estimated by the bookies? Gives us much more scope when churning out the predictions.

1) do they have a positive goal difference 2) is their mean points per game greater than the mean for the division 3) is their mean points per game for the last 4 games greater than their average for the season 4) do they have more than 50% of shots on target

The binary idea is sound in principle,keeps it simple.The bullet points are sensible, but I would question number 4 as being irrelevant, and also time consuming in the gathering in of all the data.I feel its like corners, you could get 100 and not have any goals to show for the attacking, and again I feel that this contribution would provide little in the final outcome.Lots of good stuff there Paul, keep em coming!;) btw concentrating on getting match results by means of inputting goal differences along with 7 other variables out of my NN.Will reveal some of these later after further discussions.;)

[merlin]

Re: Neural Networks as a means of prediction.

You say you are not looking for bookies odds but if you are using it as a betting tool whether directly or indirectly you are by default.

Thanks again Oddsfellow, feel as if I am like an agony aunt trying to keep up with all responses.In loose terms I suppose I am using bookies decimalised odds within the network, so on that point I yield.However,as the NN does not recognise the input as such, it only reads these in conjunction with all other input, eight variables in all in my current model, and churns out a decimal figure in the output column that has no bearing whatsoever to any odds figures quoted by the bookies.As I said previously,the only time I look at these is outside of the framework of the NN and as a means of tracking prices.The figure presented in my final column, being a decimalised figure which I can interpret as being home,draw or away.Maybe we can now put this baby to bed.By all means keep your debates coming, good for getting these points ironed out,and good for the readers.Cheers young fella.;)

[merlin]

Re: Neural Networks as a means of prediction. Sorry, nearly missed your post thanatos, and welcome aboard to you too, the more the merrier I say.I have not yet seen any Neural Networks that give you names and features as such.All data input is by way of numbers or decimals and is represented in this way throughout the NN learning process.If you have seen anywhere where'words' can be entered, let me know,would indeed be interesting. Thanks for your contribution.;)

[PAULM03]

Re: Neural Networks as a means of prediction.

btw concentrating on getting match results by means of inputting goal differences along with 7 other variables out of my NN.Will reveal some of these later after further discussions.

To be honest merlin i think it's time for cards on the table. We're all supposed to be working towards a common aim and stimulating some ideas on the use of NN through constructive criticism. That can't happen if you're holding back.

[merlin]

Re: Neural Networks as a means of prediction. I hardly think that it is cards on the table time PaulM03, I am not holding anything back of major importance.I am no expert hence the raising of the thread to stimulate some interest.Some of my ideas may be no better than yours as I am still searching for betterment all the time.I would like to see some ideas on the number of variables that we think should be used.There has been no discussion on this as yet.Also,way back I set out some criteria by way of just how we should go about setting the data out,home data to one side,away data to the other.Seen no mention of this yet.Your post went the nearest to what we are looking for, but hey, we are out to stimulate interest, and it would be great if other ideas came into the scenario before we commence to discuss my model.Any takers on this?;)

[Dino]

Re: Neural Networks as a means of prediction. Hi Merlin, if I understand things correctly,you're sticking around half a dozen variables into a NN to obtain an output figure for each of the two teams in a match up. You're then making a subjective judgement as to the result of the game.Presumably a large advantage for one team indicates a win for that team & little difference between the output indicates the likelyhood of a draw. You're problem as I see it is that if your NN is any good,your selections could just as easily be made by looking at the odds.Chelsea v WBA is going to indicate a win for the Blues,CP v Bolton is going to be closer and a draw possibility. What you need to do(via regression perhaps) is stick a home win/draw/away win % onto a match given your outputs for each team taking part in that match.Only then can you compare your odds to the available odds & that's a fundamental of any selection process. As a unrealistic example,Chelsea v WBA,Chelsea are obviously the most likely winners(say around 1/10 with WBA 20/1),but if odds of say 200/1 are available on WBA then WBA would most likely be the value bet.In real life the differences of course would be much less noticable,so selection by inspection wouldn't be possible. You need to know what a 10.7-9.6 match up "means". Then you need to be sure that your 10.7-9.6 figures have sufficient depth of data to be representative of the true longterm prospects of each team.In other words that you haven't "tossed the coin" 5 times,got 4 heads ad 1 tail and deemed it unfair....on insufficient evidence. As regards to your inputs,I plump for likely goals scored against an average opponent & likely goals conceded.It's the best correlation for future results,any other factor is less strongly correlated and almost certainly not independent of goals scored/goals conceded.

[spreadman]

Re: Neural Networks as a means of prediction. Merlin, it looks like you may be stuck between a rock and a hard place, or to continue my OK Coral, 'High Noon' imagery, it appears you stand between the gunsights of the non nn's and the nn users. Should you come out shooting or make your sherrif's speech? Or am I talking a load of 'bullet'? Make your move, dear Merlin! Good luck, Sixgun Spreadman . :ok

[merlin]

Re: Neural Networks as a means of prediction. Thanks Dino for a well structured post.Your assumptions on the whole are largely correct with the final output figures viewed to see what category the predictions are in, i.e. home,draw,win.I know absolutely zero about regression, so I don't want to get into unknown territory.The decimalised bookies odds already go into the data along with all the other variables within the learning process, and later I will be given examples of some of the data churned out.I would go along with your choices on the variables you mention in the closing paragraph, but will be looking for a two headed coin in future.:D

[merlin]

Re: Neural Networks as a means of prediction. Spreadman, I see no guns! Mind you I have the telescope up to my glass eye!I make my own decisions when it comes to the content of this thread, and I will certainly not be pressurised into jumping in with both feet.I never visualised when I started this thread that there would be attempts to put the 'heavy gang' in.It has been a difficult process in trying to explain various processes whilst employing NN methods, especially when one's NN knowledge is not what it could be,top notch.Coupled with this there does not appear to be others who could fill the bill and take up the yellow jersey,so meanwhile I soldier on and do what I consider to be best for the thread and for the interest of those trying to follow.I believe as originator of the thread, this is my right, and I will not be pushed into a corner.My move has already taken place young fella, I'm afraid it's the Sheriff's speech,High Noon or no High Noon!;) ps Thanks for the input.

[brookes]

Re: Neural Networks as a means of prediction. I was a little confused before. I then downloaded easyNN , now I am really confused. I've got no idea what to do ! How do you input the parameters ? It is like trying to fly before I can crawl.

[merlin]

Re: Neural Networks as a means of prediction. Hi there Brookes,I understand how you feel,but all is not lost.Click on the help files and tha basic excercises will get you started.I believe from memory this is one about colours, but it's introduction is user friendly and will point you in the right direction.The tutorials will also help you get the feeling of the input.Cheers.;)

[Osesame]

Re: Neural Networks as a means of prediction. If it was me 1) Separate home and away performance. 2) use a corresponding period of 2 months for previous data input 3) select inputs carefully, my selections not in order of importance would be a) goals for and against b) games won,drawn,lost c) day of the week d) converted set plays Regarding injuries suspensions etc, I think that an assessment can be made after the expected result or scoreline has been evaluated by NN.

[merlin]

Re: Neural Networks as a means of prediction. Well done, Osesame, first one in to get his feet wet!Good constructive thinking Sir!

a) goals for and against b) games won,drawn,lost c) day of the week d) converted set plays

Posted by Osesame. a) The only thing I have against this is that total goals F and A can get large and cumbersome,how would you consider average goals F & A? b) Looks acceptable. c) Day of the week? Does this have any significant value in predicting? d) Converted set plays - Where do we gather this info,and would this not be too time consuming. I am thinking in my terms here,where I usually predict all 49 coupon matches. I particularly like the idea of the after prediction scenario to find your further match news etc. Good line of thinking here Osesame, getting towards a logical layout.Cheers.:ok

[PAULM03]

Re: Neural Networks as a means of prediction. I think we've already arrived at logical layout. There are only a limited number of variables for which data is readily available to train the network and a limited number of ways in which it can be used. Osesame has it almost all covered. The problem is what do you want you neural network to tell you? If you simply want it to tell you which side is better, as your current one appears to do merlin then that seems to be a waste of computing power. Simply looking at the bookies odds will tell you the same thing, your NN doesn't bring you any closer to an objective assesment on whether a particular bet is 'value'. The only way your system will make you money long term is if it can identify matches where the probability of a particular outcome is greater than that suggested by the bookie. I've worked out how to use easyNN to predict the outcome of a particlar match. I have my variables in mind and will be subscribing to joe's site to get the historical data to 'train' the NN. I need to do a bit more thinking about converting the NN's output to a probability. First stop is Mr.O's excellent thread. Then hopefully some NN forums on the internet.

[Osesame]

Re: Neural Networks as a means of prediction. Re goals for and against. If you use averages there will be very little difference after each event, the goals for and against will help to Identify current high/low scoring teams.It is a much more variable "variable" if you see what I mean. Re converted set plays,I have no idea where this info might be, perhaps you could find some alternative.I don't like the idea of just using corners as unless some of them are converted then they do not affect the outcome.A team could have very few corners and convert 50%,another may get lots of corners but only convert 5%. Re day of the week,The only answer I can give is yes and it would be top of my personal list.There have been several discussions on this previously.To a large extent it takes in the time of day for weekday matches, but does not explain the diferences between Sat/Sun games,perhaps it is the roast beef and yorkshire pudding for lunch :loon . I note that you do all the leagues, and would strongly advise to do them separately. You would surely get some strange results if you lumped together two contrasting leagues such as Spain segunda and Eng. premier. The disimilarities between the English leagues may not be so apparent but I can assure you they exist.

[Osesame]

Re: Neural Networks as a means of prediction. Or you could work out how to find your answers at Tip-ex

[merlin]

Re: Neural Networks as a means of prediction. Thanks Osesame.Personally I don't think that I will ever use set plays as I don't see this kind of variable affecting the workings of the NN to a great extent.So I will save myself time in having to look for the info.I don't do all the leagues as you state in your post.What I try to do on a weekly basis is to input the info in relation to the variables I am using and thus try to predict the course of the 49 British Pools matches, so Spain, Italy etc are not included in the run.As for the day of the week, and you quite rightly mention roast beef and yorkshire for Sunday lunch.Might well depend on the portions lashed out to the players, who knows!Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction.

The problem is what do you want you neural network to tell you?

posted by PaulM03 Pleased you mentioned this Paul,it is the very root of the problem. Once this is determined,then the variables can be slotted in to suit your needs. I never search with a view to seeing just which side is better,would be a waste of excercise.I'm trying to look a little further than that, sometimes at the margin of victory/defeat it defines.Also I did mention earlier that we may have found some value bets by opposing the long odds home and away shots and going completely against the NN prediction.Still in the experimentation stages though.Looks like you have got yourself into a position to also experiment, and I trust that you will find it an enjoyable experience.Cheers.;)

[PAULM03]

Re: Neural Networks as a means of prediction.

Also I did mention earlier that we may have found some value bets by opposing the long odds home and away shots and going completely against the NN prediction.

1) How have you determined the value? 2) How have you decided which ones to go against? 3) It doesn't take a NN to work out it may be worth opposing teams that are at long odds, especially away.

[merlin]

Re: Neural Networks as a means of prediction.

It doesn't take a NN to work out it may be worth opposing teams that are at long odds, especially away.

You keep telling me this, don't you think that I already know!:(

[PAULM03]

Re: Neural Networks as a means of prediction. But that's what you've said you are doing:unsure and you still haven't said how you're defining value or how you're choosing which teams to oppose.

[merlin]

Re: Neural Networks as a means of prediction. .... errr No! Show me where I have said that I do not need an NN for determining which teams to oppose. :unsure As for those teams to oppose I have already mentioned this previously.The final output column gives me a reading, which is then determined as 1,2 or X. By reading this in conjunction with some of the longest priced odds either at home or away,the teams to oppose are then highlighted.It has always been the intent of this thread to endeavour to promote the use of Neural Networks for prediction with a view to attracting constructive discussion and ideas, and I will always uphold that view.;)

[PAULM03]

Re: Neural Networks as a means of prediction.

.... errr No! Show me where I have said that. I do not need an NN for determining which teams to oppose.

Also I did mention earlier that we may have found some value bets by opposing the long odds home and away shots and going completely against the NN prediction

Am i misunderstanding what you're trying to get across?