Predicting the result of a Virtual football match using tidymodels
After successfully scraping data from a football betting site i will now build a simple model using tidymodels r package to see how well it will perform in predicting the result of the games.
Below are some information about the dataset:
- home –> A premier league club at home to play against an opponent
- away –> A premier league club away from home to play against the home team
- homeOdd –> the odds that the home team will win a particular match
- drawOdd –> the odds that both teams will draw a particular match
- awaysOdd –> the odds that the away team will win particular match
- homeR –> result of the home team after a particular match
- awayR –> result of the away team after a particular match
- matchtime –> the time a particular match was played
- leagueW –> the week a particular match was played
- day –> the day a particular match was played
install the tidymodels r package
install.packages("tidymodels") # skip this step if you already have the tidymodels package
load the package
require(tidyverse)
Read in the dataset and check its dimension
bet9ja <- read_csv("https://raw.githubusercontent.com/twirelex/dataset/master/bet9ja(2).csv")
## Parsed with column specification:
## cols(
## home = col_character(),
## away = col_character(),
## homeOdd = col_double(),
## drawOdd = col_double(),
## awaysOdd = col_double(),
## homeR = col_double(),
## awayR = col_double(),
## matchtime = col_time(format = ""),
## leagueWK = col_character(),
## day = col_character()
## )
dim(bet9ja) ## see the dimension of the data
## [1] 14410 10
Overview of the data
require(knitr)
kable(head(bet9ja))
| home | away | homeOdd | drawOdd | awaysOdd | homeR | awayR | matchtime | leagueWK | day |
|---|---|---|---|---|---|---|---|---|---|
| WAT | LIV | 6.08 | 4.37 | 1.52 | 0 | 3 | 16:32:00 | WEEK 34 | Friday |
| EVE | NWC | 2.38 | 3.20 | 3.13 | 0 | 2 | 16:32:00 | WEEK 34 | Friday |
| TOT | BUR | 1.46 | 4.38 | 7.18 | 3 | 1 | 16:32:00 | WEEK 34 | Friday |
| BOU | MNU | 3.54 | 3.00 | 2.30 | 0 | 1 | 16:32:00 | WEEK 34 | Friday |
| BRI | ASV | 2.29 | 3.02 | 3.53 | 0 | 1 | 16:32:00 | WEEK 34 | Friday |
| CHE | NOR | 1.46 | 4.38 | 7.18 | 3 | 1 | 16:32:00 | WEEK 34 | Friday |
glimpse(bet9ja)
## Rows: 14,410
## Columns: 10
## $ home <chr> "WAT", "EVE", "TOT", "BOU", "BRI", "CHE", "WHU", "WOL", "...
## $ away <chr> "LIV", "NWC", "BUR", "MNU", "ASV", "NOR", "SOU", "MNC", "...
## $ homeOdd <dbl> 6.08, 2.38, 1.46, 3.54, 2.29, 1.46, 1.94, 4.16, 3.02, 2.0...
## $ drawOdd <dbl> 4.37, 3.20, 4.38, 3.00, 3.02, 4.38, 3.32, 3.81, 3.19, 3.2...
## $ awaysOdd <dbl> 1.52, 3.13, 7.18, 2.30, 3.53, 7.18, 4.25, 1.82, 2.46, 3.9...
## $ homeR <dbl> 0, 0, 3, 0, 0, 3, 1, 2, 1, 1, 2, 1, 0, 3, 3, 5, 1, 0, 2, ...
## $ awayR <dbl> 3, 2, 1, 1, 1, 1, 0, 1, 1, 1, 0, 0, 2, 0, 3, 1, 3, 3, 1, ...
## $ matchtime <time> 16:32:00, 16:32:00, 16:32:00, 16:32:00, 16:32:00, 16:32:...
## $ leagueWK <chr> "WEEK 34", "WEEK 34", "WEEK 34", "WEEK 34", "WEEK 34", "W...
## $ day <chr> "Friday", "Friday", "Friday", "Friday", "Friday", "Friday...
we do not have a response variable yet so we need to first create a response variable before proceeding.
we will create the new variable from our homeR and awayR variables and call the new variable “result”, we will encode the “result” variable as follow:
- 1 when the home team wins
- 0 when the game ends in a draw
- 2 when the away team wins
bet9ja <- bet9ja %>%
mutate(result = factor(case_when(homeR > awayR ~ 1, homeR < awayR ~ 2, TRUE ~ 0))) %>%
select(-homeR, -awayR) # remove the homeR variable and awayR variable as we will no longer need them
we have now added a 3 category response/dependent variable and this means that we will be building a multiclass classification model.
some of the variables are represented as character variables but we will need them to be categorical variables for the sake of the algorithm we want to make use of.
bet9ja <- bet9ja %>%
mutate_if(is.character, as.factor)
DATA EXPLORATION
If you are familiar with the way football betting works you will know that “odds” play a huge role on whether a team will win a particular game or not. So let us first visually see the relationship between the odds variables and the result variable.
Let us discretize the odds variables and do a plot
bet9ja %>%
transmute(result, homeOdd = factor(
case_when(
homeOdd <= 1.5 ~ "smallO",
homeOdd > 1.5 &
homeOdd <= 2 ~ "mediumO",
homeOdd > 2 ~ "bigO"
)
)) %>%
ggplot(aes(result, fill = result)) + geom_bar(show.legend = FALSE) + facet_wrap( ~ homeOdd) + theme_bw() + scale_y_continuous(labels = NULL) + labs(y = NULL, title = "HOME TEAM ODDS PLOT") + theme(plot.title = element_text(hjust = 0.5))
bet9ja %>%
transmute(result, awaysOdd = factor(
case_when(
awaysOdd <= 1.5 ~ "smallO",
awaysOdd > 1.5 &
awaysOdd <= 2 ~ "mediumO",
awaysOdd > 2 ~ "bigO"
)
)) %>%
ggplot(aes(result, fill = result)) + geom_bar(show.legend = FALSE) + facet_wrap( ~ awaysOdd) + theme_bw() + scale_y_continuous(labels = NULL) + labs(y = NULL, title = "AWAY TEAM ODDS PLOT") + theme(plot.title = element_text(hjust = 0.5))
MODELING
Now that we have an idea of what the relationship between the odds variables and the result variable looks like let us go ahead and build a predictive model using the tidymodels r package.
let’s build a svm model using tidymodels parsnip
require(doSNOW) # for parallel processing in windows OS
cl <- makeCluster(6, type = "SOCK")
registerDoSNOW(cl)
split data into training and testing
require(tidymodels) # for modeling
set.seed(111) # setting seed for reproducibility
splitdata <- initial_split(bet9ja)
traindata <- training(splitdata)
testdata <- testing(splitdata)
fitting
svm_model <- svm_rbf(mode = "classification") %>%
set_engine("kernlab") %>%
fit(result~., data = traindata)
make prediction on test data and view confusion matrix
prediction <- svm_model %>%
predict(testdata) %>%
transmute(actual = testdata$result, rename(.,predicted = .pred_class))
kable(head(prediction, 10))
| actual | predicted |
|---|---|
| 2 | 1 |
| 1 | 1 |
| 2 | 2 |
| 1 | 1 |
| 0 | 1 |
| 1 | 1 |
| 1 | 2 |
| 1 | 1 |
| 0 | 2 |
| 0 | 1 |
| we can see that there are few misclassifications |
prediction %>%
conf_mat(actual, predicted) # view confusion matrix
## Truth
## Prediction 0 1 2
## 0 5 10 12
## 1 720 1253 540
## 2 306 246 510
prediction %>% accuracy(actual, predicted) # check accuracy
## # A tibble: 1 x 3
## .metric .estimator .estimate
## <chr> <chr> <dbl>
## 1 accuracy multiclass 0.491
The svm model produced an accuracy of about 49%
Now let us try to build a random forest model to see if there will be any improvement
rf_model <- rand_forest(mode = "classification") %>%
set_engine("ranger", importance = "permutation") %>%
fit(result~., data = traindata)
make prediction on test data and view confusion matrix
prediction2 <- rf_model %>%
predict(testdata) %>%
transmute(actual = testdata$result, rename(., predicted = .pred_class))
kable(head(prediction2, 10))
| actual | predicted |
|---|---|
| 2 | 1 |
| 1 | 1 |
| 2 | 2 |
| 1 | 1 |
| 0 | 2 |
| 1 | 1 |
| 1 | 2 |
| 1 | 1 |
| 0 | 2 |
| 0 | 0 |
prediction2 %>%
conf_mat(actual, predicted) # view confusion matrix
## Truth
## Prediction 0 1 2
## 0 139 146 151
## 1 583 1090 441
## 2 309 273 470
prediction2 %>% accuracy(actual, predicted) # check accuracy
## # A tibble: 1 x 3
## .metric .estimator .estimate
## <chr> <chr> <dbl>
## 1 accuracy multiclass 0.472
The svm model performed better judging with the Accuracy
lets view how well each of the variables performed in the model
require(vip) # variable importance plot function "vip()"
rf_model %>%
vip(geom = "point")
OBSERVATION
Both models didn’t give us the best result but from our analysis we can see that variables that are the most significant determinant of the outcome of any virtual football game are the odds variables, the teams involved or the time/day the game is played doesn’t matter that much…
CONCLUSION
With an accuracy less than 50% the SVM and a random forest model are probably not the best models to depend on when it comes to virtual football game prediction.
Wrap-up
Shifting from the base r and caret way of modeling can be hard for some of us but seeing how far the tidymodels is preparing to take us (timely upgrade/update) is enough reason for everyone to start trying it out. And also, knowing that the brain behind the caret package is also part of the tidymodels team makes it even more interesting. I have prepared a short and comprehensive lesson on tidymodels that might be helpful.. Tidymodels for beginners
Anthony Aigboje Akhonokhue
A statistician/data-analyst/data-scientist for hire. See my Resume for more info.