## Sunday, September 9, 2012

### Football predictions display

Having looked at the football data earlier, I wanted to look at predictions for new games. This consists of two parts, getting a predictive model, predicting and displaying the predictions. I decided to do this backwards, first to make the displays. This will make things easier when the time is there to compare models. To get the predictions I use a very simple model, which basically states, a club makes about x goals, irrespective of all other conditions. I don't believe this model, but it can give predictions.
model1 <- glm(Goals ~OffenseClub,data=StartData,family='poisson')
The consequence of this setup is that each game needs two predictions, one for the first club, one for the second. For clubs Vitesse and FC Groningen are used to make the predictions.

The prediction of the glm is a mean number of goals, which is still quite far from the reality of a number of goals. For this I use the Poisson distribution and treat the prediction as true. I do not include overdispersion nor standard error of parameters. The result shows FC Groningen has 30% of getting no goals, 35% chance of getting 1 goal, 22 % for two goals, after which the chances become quickly very low.
top <- data.frame(OffenseClub=c('FC Groningen','Vitesse'))
prepred <- predict(model1,top,type='response')
(dp1 <- dpois(0:5,prepred[1]))[1:4]
[1] 0.29942768 0.36107456 0.21770672 0.08750956
Finally, the predictions need to be combined, to get a pair of goals. Not surprisingly, if the chance of a particular outcome, such as one goal, is 30%, then the chance of a pair of outcomes, such as 1-1 may be 30%*30%=9%. In this case it turns out to be slightly higher, 12%. This is the most probable outcome too.
dp2 <- dpois(0:6,prepred[2])
oo <- outer(dp1,dp2)
rownames(oo) <- 0:6
colnames(oo) <- 0:6
round(oo,digits=3)
0     1     2     3     4     5     6
0 0.073 0.103 0.073 0.034 0.012 0.003 0.001
1 0.088 0.124 0.088 0.041 0.015 0.004 0.001
2 0.053 0.075 0.053 0.025 0.009 0.002 0.001
3 0.021 0.030 0.021 0.010 0.004 0.001 0.000
4 0.006 0.009 0.006 0.003 0.001 0.000 0.000
5 0.002 0.002 0.002 0.001 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000
It will be more useful to summarize the outcome as win-equal-lost. These are extracted as sums of probabilities.
c(sum(oo[upper.tri(oo)]),sum(diag(oo)),sum(oo[lower.tri(oo)]))
[1] 0.4167411 0.2612236 0.3211237
It is practical to fit all this in a little function which creates these data in one go. The only new things are the introduction of a new class fboo which is used to direct the prediction to the appropriate accompanying print function and some attributes to administrate the clubs predicted.
fbpredict <- function(object,club1,club2) {
top <- data.frame(OffenseClub=c(club1,club2),DefenseClub=c(club2,club1),OffThuis=c(1,0))
prepred <- predict(object,top,type='response')
dp1 <- dpois(0:9,prepred[1])
dp2 <- dpois(0:9,prepred[2])
oo <- outer(dp2,dp1)
rownames(oo) <- 0:9
colnames(oo) <- 0:9
class(oo) <- c('fboo',class(oo))
attr(oo,'row') <- club1
attr(oo,'col') <- club2
wel <- c(sum(oo[upper.tri(oo)]),sum(diag(oo)),sum(oo[lower.tri(oo)]))
names(wel) <- c(club1,'equal',club2)
return(list(details=oo,'summary chances'=wel))
}

print.fboo <- function(x,...) {
cat(attr(x,'row'),'in rows against',attr(x,'col'),'in columns \n')
class(x) <- class(x)[-1]
attr(x,'row') <- NULL
attr(x,'col') <- NULL
oo <- formatC(x,format='f',width=4) # fixed format
oo <- gsub('\\.0+\$','       ',oo)   # replace trailing 0 by ' '
oo <- substr(oo,1,6)                # and fix the width
print(oo,quote=FALSE,justify='left')
}

fbpredict(model1,'FC Groningen','Vitesse')
\$details
FC Groningen in rows against Vitesse in columns
0      1      2      3      4      5      6      7      8      9
0 0.0730 0.0880 0.0531 0.0213 0.0064 0.0016 0.0003 0.0001 0      0
1 0.1030 0.1242 0.0749 0.0301 0.0091 0.0022 0.0004 0.0001 0      0
2 0.0727 0.0877 0.0529 0.0213 0.0064 0.0015 0.0003 0.0001 0      0
3 0.0342 0.0413 0.0249 0.0100 0.0030 0.0007 0.0001 0      0      0
4 0.0121 0.0146 0.0088 0.0035 0.0011 0.0003 0.0001 0      0      0
5 0.0034 0.0041 0.0025 0.0010 0.0003 0.0001 0      0      0      0
6 0.0008 0.0010 0.0006 0.0002 0.0001 0      0      0      0      0
7 0.0002 0.0002 0.0001 0      0      0      0      0      0      0
8 0      0      0      0      0      0      0      0      0      0
9 0      0      0      0      0      0      0      0      0      0

\$`summary chances`
FC Groningen        equal      Vitesse
0.3213815    0.2612237    0.4173918

1. Nice!
"Having looked at the football data earlier..." where it was posted before?
Thanks, Fernando

1. Hi Fernando, on 28 August I collected the data of last year, reshaped it and looked if the counts confirm to a Poisson distribution. Please see this link http://wiekvoet.blogspot.nl/2012/08/football-eredivisie-goals.html

2. When you watch the University of Georgia football games online, there is one undeniable image totally inescapable - the Bulldog mascot. It is perhaps one of the most easily recognizable college football mascots of all time. And, choice of the bulldog as the image for Georgia sports was due to a desire to associate the teams with the fierceness this particular animal portrays.

Football bingo
Grassroots football

3. Hi there,
I ran into your document on the internet, I couldn't understand much of it but what caught my interest is your believability that it is possible to predict football match results....
I will like to know if you have  little time for me to share what I have with you...
There is another way to predict football draws but there is no way to document it because academia isn't my strong forte....

I would like to know if you are conversant with R or Weka or Rapid Miner?

My method of prediction is crude and I would like to know if you and I cab brainstorm on how to engage machine learning techniques to increase chances if predicting the right football match result....
If this discussion interests you, please get back to me.

Iamshuga@gmail.com