Neural Networks as a means of prediction.

[merlin]

Re: Neural Networks as a means of prediction. I can see just where you are coming from now.Yes, you have completely misunderstood Paul.

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

Posted by me originally. Reading the last part of the quote...... and going completely against the NN prediction. At this point the NN has churned out the predictions,as being 1,2 or X then the latter action of finding the teams to oppose takes place.There has never been any mention of non usage of the Neural Network as a model.Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction. ...... OK looks a good time to introduce those variables I chose in my last model. So it's Christians to the Lions time eh!:D All variables as far as possible reading from left to right.Seven intput columns and column 8 is the output column. 1) Bill Hunter's Yearbook Ratings.Home team - Away Team =Yb difference. 2) Racing & Football Outlook Index Ratings - Home - Away = RFO difference. 3) Ladbrokes Home decimal odds. 4) Ladbrokes Draw decimal odds. 5) Ladbrokes Away decimal odds. 6) Home Trend taken from RFO Index ratings. 7) Away Trend taken from RFO Index ratings. 8) Goal difference of rated match. I'ts not proved to be a winner in terms of good percentage correct predictions but has possibilities in opposing home/away predictions at longer odds. So it's onwards and upwards!Cheers.;)

[mr onemore]

Re: Neural Networks as a means of prediction. Hi Merlin, A brave man indeed - posting these for us all to discuss!!:D Can I just check the mechanics of these NN's then. Am I right in thinking the NN takes the 7 inout variables above - and plays around with them in optimal ways to try and match the output column - the goal difference that actually occurred? If that's the gist - then I really would like to see you try your Network again, but with cols 3,4,5 left out. It would be interesting to see how the new results would compare with your old ones. Interested to see what happens here...

[merlin]

Re: Neural Networks as a means of prediction. Hi there Mr.O. Greetings from my deeply dug trench, wearing my tin hat at the moment with rifle and bullets at the ready! The whole 8 variables are entered into the NN from past matchups including results.The NN is then trained in learning these as a set of variables and data relating to each match.The NN model I have at the moment has a genetic net and learns over several generations.Then produces what it considers to be predictions based on those variables etc.When I have say a full coupon list of 49 matches to predict,the data is then entered minus column 8 as the matches have not yet taken place.The NN then produces it's list of predictions.An interesting point you have there about leaving out cols 3,4 and 5, will give that a try and build a new network with a view to testing out this new theory.Cheers.;)

[merlin]

Re: Neural Networks as a means of prediction. ......... hmmmmm! Does not seem to be as many as I thought that may pass comments and/or judgement on those variables.They must make some sense after all!;)

[dietcoke]

Re: Neural Networks as a means of prediction.

...... All variables as far as possible reading from left to right.Seven intput columns and column 8 is the output column. 1) Bill Hunter's Yearbook Ratings.Home team - Away Team =Yb difference. 2) Racing & Football Outlook Index Ratings - Home - Away = RFO difference. 3) Ladbrokes Home decimal odds. 4) Ladbrokes Draw decimal odds. 5) Ladbrokes Away decimal odds. 6) Home Trend taken from RFO Index ratings. 7) Away Trend taken from RFO Index ratings. 8) Goal difference of rated match.

To me 1) and 2) are essentially the same. They will certainly be highly correlated. 3),4),5) Theres been a lot of discussion about use of bookie odds and I fully agree with the consensus that they should not be part of the input. 6) 7) My attitude to these depends on what RFO figure you are talking about. Do you mean the RFO trend figure for both home and away teams or, do you mean the RFO home only and away only ratings?? If its the former, you are not as far as I can see incorporating any home advantage into your network, if its the latter, you have no recent form indicator. How much RFO data have you got? I'd be inclined to generate some predictions using just that. Using that, you then have the ability to compare your predicted supremacies against RFO prediced odds and supremacies.

[merlin]

Re: Neural Networks as a means of prediction. Hi there Dietcoke, just seen your message,and thanks for your input and further thoughts, and better still, can I take my tin hat off yet, it's kinda getting warm underneath this lot.:D Variables 1 and 2 were used as to provide consencus of opinion as to the relative strengths of the two sides.By taking both ratings it was felt that this would then be achieved.OK I have taken on board the comments relating to variables 3,4 and 5 and tend to agree with Onemore who first raised this point,that I will try a further network without using these.The trend figures used were the figure in the extreme right column of the Index Ratings.There is only one figure against each team unfortunately, but it was included because it was thought that this would give weight as to the way each team was performing.Again this period of time is unknown due to the fact that the RFO do not reveal how the trend figures are worked out,but as some measure of current form needed to be included, we used the only figure we had.I still feel strongly that there is a case for some form of average opponents strength figure ought to be included,in order to give a clearer picture of the relative strengths of teams played so far to date.It would be a much clearer indication than using the league Tables I feel.In your closing paragraph you mention using home advantage, but again I feel that there will be home advantage already built into RFO Home Ratings of some description,although this is uncertain due to the unknown formation and calculation of such information.Thanks for your constructive comments.;)

[merlin]

Re: Neural Networks as a means of prediction. Any further comments to make on post #127 Dietcoke? Or for that matter does anyone else have any further thoughts/ideas on Neural Networks?Have we any further input of variables in particular, as I feel that we have just scratched the surface.Cheers.Merlin.;)

[snyena]

Re: Neural Networks as a means of prediction. Distance the away team has to travel, especially when they dont fly, in the lower divisions, and when they travel not too long before the matchday. There was a thread on this last year. I am interested in finding out whether this variable could have a substantial impact on the outcome of a game and not be reflected in the odds accurately. I am not sure whether it is suitable for Neural Networks though.

[dietcoke]

Re: Neural Networks as a means of prediction. There is certainly some effect. I found a research paper on the web a while back which used this feature a one of number of inputs to a statistical based system. The guy who did the research found a 10% correlation with the result. More than shots on target and corners. This could be used in an ANN but would be best as a categorical variable. where milages are grouped into low,medium and high(1,2,3). My view is that distance only really has an effect over a certain milage

[merlin]

Re: Neural Networks as a means of prediction. Certainly a valid point raised by you two gents, and one to bear in mind for the future.As you say, probably has more effect on the lower divisions than up above.Be interesting to see how this affects NN calculations.Also I have it in the back of my mind to try out average opponents strengths,seems to me to make better sense than merely using the league tables,as it is fairly representative of the strength of those teams played to date.Food for thought?;)

[merlin]

Re: Neural Networks as a means of prediction. Where is PaulM03, I thought that you were preparing yourself to do some experimenting, not heard anything for a while, how's it going?;)

[merlin]

Re: Neural Networks as a means of prediction. I'm just wondering whether there is a case for Major League Baseball to be introduced to the Neural Networks scenario.There appears to be more than ample statistics around a number of American sites, which would lend itself very nicely to the data required.Infact they tend to go overboard with all the sites and stats and pitcher matchup etc, so may look into this with a view to doing something a little different.;)