A look at our latest challenge: The Cricket World Cup Challenge
Context
Here in the office, we’re 50% cricket mad (the others not so much), so we wanted to put on a fun competition to do with the cricket world cup. We reached out to our network to see if anyone had any ideas. What came back is really exciting, and we can’t wait to see what you all come up with.
Mustard Systems have kindly given you all access to their cricket dataset of over 1million data points and over 20 features. Their challenge is simple: Given all the information up to a particular ball, what is the chance that that ball will be a wicket?
To get you started, we want to show you how this problem has been tackled in the past.
Crickytics Approach
Faizan at Crickytics attempted a similar version of this problem using a smaller data set with 11 features. His approach was to turn the problem into a binary classifier on predicting whether there would be a wicket in the next 6 balls.
After attempting the classification in several ways, he found that the optimal approach was to limit the model to data from the last 18 balls (3 overs) and then use a mixture of random forest and feedforward neural networks. Interestingly he found that the explicit output of the classifier wasn’t accurate.
Instead, he found that if the model predicted no wicket with a P<0.9, then that was likely to be the case in the data. Anything lower than this (e.g. if the model made a classification of no wicket (P<0.9)), was likely to be a wicket.
By tweaking his output like this he managed to achieve the following results:
- True Positive Rate: 39%
- True Negative Rate: 87%
Have a closer look at his write-up here: https://crickytics.com/predicting-next-wicket-in-a-t20-match-using-ai/
Competition Registration
Registrations open on the 25th May and the competition starts on the 1st June. Once you’ve registered you will be sent the feature headings so you can start thinking about how you might tackle the problem!
