The prizes in snooker can be sometimes ridiculously huge so why don't we use data science to predict some?

Snooker Prizes

This project's goal is to predict snooker players' money prizes in whole season basing on his achievements.

It's a web application, which downloads data from online resources and perform predictions.

Web scraping

The code uses Python's requests to fetch data and BeautifulSoup to scrap it - downloaded HTML code is being searched for particular blocks (e.g. find all div tags of class abc123).

Pros:

It's very fast

Cons:

It's very vulnerable for any format changes on the website (for instance changes in page layout or renamed tag classes).

To prevent this vulnerability at least a bit, all links are configurable as a part of CueTracker.py in Links class:

class Links(Enum):
    CENTURIES = "centuries/most-made/"
    TITLES = "tournaments/won/"
    TOURNAMENTS_AND_MATCHES = "matches-and-frames/won/"
    MONEY = "prize-money/won/"

Data

All of data is being downloaded from https://cuetracker.net - a website with loads of snooker data. This is done dynamically, whenever the program is running. This could be optimized by downloading the data every interval (e.g. every 1 week or 1 month) and save it on disc space.

Since the data could be heavy (e.g. whole HTML code) some of the variables are deleted from memory right after their usage.

Predictions

To predict money prizes a Random Forest model has been used.

Random forest model is a combination of multiple decision trees. The algorithm randomly chooses features, randomly chooses sample from training set and builds a decision tree. Then it repeats the process multiple times. When all trees (i.e. a forest) are created, the model is ready. Predictions are being made by passing a piece of data through the whole forest and calculate mean from produced values.

There are few parameters that should be tuned within Random Forest model:

• max_features - the maximum number of features that will be used in every single decision tree. It affects performance: the bigger a value is, the algorithm works slower, but the results will be more accurate.
• min_samples_leaf - the minimum number of samples required to be at a leaf node. This parameter is similar to min_samples_splits.
• n_estimators - number of trees in the forest. It affects performance as well.
• bootstrap - boolean parameter to specify if bootstraping will be used (it means that the observations within one sample can occur more than once). If False, standard subsampling will be used.

The model performs quite well, since the accuracy (computed with Cross Validation) is more than 80%, despite quite small data sample .

Variables

• "Name" - player's first name and last name, could be used as an index
• "Centuries" - how many times player achieved a break with at least 100 points,
• "T_Played" - how many tournaments player has taken part in,
• "M_Played" - how many matches player has played,
• "M_Won" - how many matches player has won,
• "Titles" - how many tournaments player has won,
• "Money" - sum of all money prizes (target variable).

All values are applicable for chosen seasons.

Web app

The frontend and routing logic have been developed with Flask, which is a framework designed specifically for simple web apps.

Libraries

• BeautifulSoup - to parse and scrap HTML code.
• Pandas/NumPy - to prepare data for use in DataFrame format.
• sklearn - to perform predictory computations.
• requests - to call HTTP methods on specifil URL. Selenium solution is also possible, but since it would require much more development, requests lib has been used.

Hosting

The project is ready to be hosted using Heroku.