2
$\begingroup$

I am trying to create a machine learning model to predict the winner of an upcoming cricket tournament (winners of all matches in the tournament). I have couple of questions here:

  1. What kind of input data can I use for training? I can't use information like who won the toss, how much did the teams score in each innings, etc because I wouldn't be having those data for the final prediction dataset
  2. What kind of algorithms should I be looking at? The prediction should be having one of the two teams participating in the match. How do I tell this to the model? Or should I build a multiclass model which predicts one of the all possible teams?

Any inputs on how to proceed further will be very useful for me because I have never worked on sports based data/models earlier.

$\endgroup$
2
  • $\begingroup$ Hi and welcome to SE DS! Can you clarify whether you are trying to predict a match or a tournament winner? $\endgroup$
    – Jonathan
    Oct 6, 2021 at 10:56
  • $\begingroup$ All the winners of all matches in the tournament. I'll update the same in the question as well. $\endgroup$ Oct 6, 2021 at 15:34

2 Answers 2

1
$\begingroup$

I'll answer your second question first!

Lets take the example of IPL. For you to predict which team will win a match/tournament, you would need to build a multiclass classification model. The output of which can be one of teams mentioned in the data. For example you can have a dataset where 2 of the features are the 2 teams participating and the model will predict which one will win. There are tons of classification model you can use for this purpose a list of which is provided here. The link provides all the available algorithms for classification and regression.

Now for your first question, Yes you are right as you cannot include data like which team scored how many runs as you will not have that info available before a match so the obvious step would be to include any kind of data that might be available prior to a match.

You would need any kind of data of the particular teams participating in a match. For example for the match Chennai Superkings vs Delhi Capitals, you would need past match stats of both the teams. You would need to collect all and any kind of past stats from previous IPL's some of which can be who won the toss, who chose batting/bowling first, pitch conditions, weather conditions, weather dew is present or not, which team has the purple cap/orange cap holder etc. All these kinds of stats are available before a match starts.

$\endgroup$
1
  • $\begingroup$ Thanks for the response @spectre. $\endgroup$ Oct 7, 2021 at 6:47
1
$\begingroup$

Generally match prediction (across sports) relies strongly on feature engineering. There are 3 types of features being used:

  1. Basic features, e.g. team names, type of tournament, current team rank, odds incl. implied probabilities
  2. Lag features, e.g. previous winner of this matchup, win rates per team, days since last match per team
  3. Complex features, e.g. ELO scores, PI ratings, running scores

Which features are most important is specific to the type of sports, e.g. in football "days since last match" is quite important but less so in esports. The 3rd group has usually the highest predictive value and is crucial (besides odds by professional bet providers). Moreover, text-based features, e.g. based on social media, can have strong predictive power too.

Model-wise gradient boosted decision trees and neural networks are among the most successful models for match prediction.

I recommend reading some of the relevant papers, e.g. for cricket specifically:

Increased Prediction Accuracy in the Game of Cricket using Machine Learning

The Cricket Winner Prediction With Application Of Machine Learning And Data Analytics

Predicting The Cricket Match Outcome Using Crowd Opinions On Social Networks: A Comparative Study Of Machine Learning Methods

$\endgroup$
1
  • $\begingroup$ Thanks for the response @Sammy $\endgroup$ Oct 7, 2021 at 6:54

Not the answer you're looking for? Browse other questions tagged or ask your own question.