Dynamic cricket match outcome prediction

3.1Data mining

This is different from the earlier techniques that have been implemented in the papers. Most of the research around predicting the match outcome takes into account a lot of different features and only makes one prediction. We have tried to adopt a different approach where we make predictions on the outcome of a match when the second team starts batting. This prediction is in regards to the second batting team and if the team will win or lose based on the match circumstances. Historical matches and player data for the Twenty20 IPL competition was scraped from the archive section of cricket fan site cricinfo.com4 for the seasons 2008 to 2019. This raw data was then cleaned and combined into two datasets for matches and deliveries. From these datasets, features sets are constructed that form the input for our model using automated excel macros. With these two datasets as input we formulated rules in the excel which in turn constructed required features of our interest for all the past Twenty20 IPL matches. The features that we have been constructed to train the model are different from any other set of features that have been used before. Total 90 attributes are created by the excel macros (Fig. 2)

Fig. 1

Exploratory Data Analysis on the constructed features.

Fig. 2

Excel Macros for features constructions.

3.2Feature selection

All matches from season 2008 to 2018 are used to train our model. This train dataset comprises of 692 matches and approx. 80,000 data points. Test data set we have selected consists of 59 matches of season 2019 and approx. 7,000 data points for prediction. No one has taken before such high data set for their predictions. We have total 90 features, which when fed to the model gives poor results. Such bad performance of the model is due to “Curse of Dimensionality”. The curse of dimensionality is one of the most important problems in multivariate machine learning. The method that we have used for feature selection is wrapper method. Wrapper feature selection methods create many models with different subsets of input features and select those features that result in the best performing model according to a performance metric. These methods are unconcerned with the variable types, although they can be computationally expensive. We identified 30 attributes that best fit to our models. The feature set as shown in Table 1, comprises of our dependent and independent variables. Here “Result” is our target categorical dependent variable. Independent parameters are constructed through excel macros utilities from our available datasets of matches and deliveries. Here we have constructed 11 batsmen and 10 bowlers and all their metrics are tracked as a function of each ball of each over throughout the match during the second inning. Also the “Target” score during the second inning is set as one of the independent attribute which gets updated dynamically as the second inning batting metrics changes as we can see from the table below, these are the final set of feature that are fed to the model. Our input data is first splitted into train dataset and test dataset on the basis of season. Then we perform KFold TimeSeriesSplit on our train dataset to ensure that we have not introduced any bias–variance trade-off in our model. In statistics and machine learning, the bias–variance trade-off is the property of a set of predictive models whereby models with a lower bias in parameter estimation have a higher variance of the parameter estimates across samples, and vice versa. Final predictions are made on the test data and model performance is evaluated.

Fig. 3

Architecture of the Model (Flow Diagram).

4.1MODEL 1 (1–20 over)

For this model we considered 692 matches of all the seasons (2008–2018) to train our model and all the 59 matches of season (2019) to test it. Train size is 79,129 and test size is 6,960. The comparative classification report given out by the model for all the classical and neural algorithms are provided below in the Table 3. As we can see in the below analysis table, among traditional machine learning algorithms SVC and RandomForestClassifier show very good results. The train accuracies are 76.47 %with confidence interval of 3.77%and test accuracies are 72.74%. Both the models have very high precision of 83.25 %and 81.90%. XGBClassifier also have very high precision of 83.92%. In deep learning models one-to-one sequence vanilla LSTM shows good accuracy of 75.05%with the confidence interval of 3.12%. We can see that the recall score and F1 score of deep learning model are comparatively higher than traditional models which means deep models are better able to classify the classes correctly than traditional models. Gaussian Naïve Bayes and Bernoulli Naïve Bayes shows very less mean train and test scores. This might be due to its assumption that attributes are independent and attributes are continuous in nature. We can see below comparative bar graph (Fig. 4) of these results to have better visualization of performance of each algorithm.

Fig. 4

Graph Classification Report Results (Over 1–20).

4.2MODEL 2 (1–5 over)

For this model we considered all records from 1 to 5 over of all seasons (2008–2018) to train our model and all 1 to 5 over of seasons (2019) to test it. Train size is 21,662 and test size is 1,820. Here SVC from traditional machine learning family and one-to-one sequence vanilla LSTM from deep learning models shows good results during the start of second inning itself. Below results in Table 4 are from first 5 overs of the second inning. The train accuracies are 68.71%, and 68.89%with confidence interval of 1.90%, and 2.20%and test accuracies of 68.68%, and 68.57%respectively. Logistic Regression is showing exceptionally high precision score of 79.27%. Again below is the comparative bar graph (Fig. 5) of results to have better visualization of performance of each algorithm between 1–5 overs.

Fig. 5

Graph Classification Report Results (Over 1–5).

4.3MODEL 3 (6–10 over)

For this model we considered all records from 1 to 10 over of all seasons (2008–2018) to train our model and all 6 to 10 over of seasons (2019) to test it. Train size is 42,793 and test size is 1,820. Here also, again SVC from traditional machine learning and one-to-one sequence stacked LSTM from deep learning models shows good results during the next 5 overs of second inning. Table 5 results are from next 6 to 10 overs of the second inning. The train accuracies are 71.87%, and 72.06%with confidence interval of 2.07%, and 2.73%and test accuracies of 72.36%, and 72.09%respectively. One-to-one sequence vanilla GRU is showing exceptionally high Recall score of 82.75%. Again below is the comparative bar graph (Fig. 6) of results to have better visualization of performance of each algorithm between 6–10 overs.

Fig. 6

Graph Classification Report Results (Over 6–10).

4.4MODEL 4 (11–15 over)

For this model we considered all records from 1 to 15 over of all seasons (2008–2018) to train our model and all 11 to 15 over of seasons (2019) to test it. Train size is 63,007 and test size is 1,829. Here as we can see deep learning models are showing best test scores as compared to traditional machine learning models. Table 6 results are from next 11 to 16 overs of the second inning. One-to-one sequence vanilla GRU and One-to-one Sequence Bidirectional LSTM are showing train accuracies of 76.02%and 75.58%with confidence interval of 3.50%, and 3.41%respectively. They are also showing good Precision, Recall and F1 scores near around 82.00%. As we can also see RandomForestClassifier is giving us a very good precision score of 90.26%on the test data. Again below (Fig. 7) is the comparative bar graph of above results to have better visualization of performance of each algorithm between 11–15 overs.

Fig. 7

Graph Classification Report Results (Over 11–15).

4.5MODEL 5 (16–20 over)

For this model we considered all records from 1 to 20 over of all seasons (2008–2018) to train our model and all 16 to 20 over of seasons (2019) to test it. Train size is 79,129 and test size is 1,491. Logistic Regression shows accuracy of 76.48%(+/–2.94) during the last part of the match which is good. SVC and one-to-one sequence of bidirectional GRU also gives accuracies close to 76.00%. They are giving test score of 77.26%and 77.13%respectively as we are close to end of the second inning. We can see that almost all models are showing a very good Recall

score of close to 100.00%which is very good as it shows in low false negative rate. So our model had ability to find all the relevant cases within the dataset. While recall expresses the ability to find all relevant instances in a dataset, precision expresses the proportion of the data points our model says was relevant actually were relevant. Below is the comparative bar graph (Fig. 8) of above results to have better visualization of performance of each algorithm between 16–20 overs.

Fig. 8

Graph Classification Report Results (Over 16–20).

Below (Fig. 9) is the comparative trend graph of train and test scores for all the traditional and deep learning models during the second inning of the match. The blue line shows the train score and the green line shows the test score across various intervals of overs. One thing we can observe across the models is that almost all the models give very good prediction during the start of 15th over. Test scores are higher than train scores. But still there are few models like SVC, XGBClassifier, and one-to-one stacked GRU which gives good test scores as compared to their train score from start of the second inning itself.

Fig. 9

Trend chart of Train and Test scores during the second inning.