In the realm of StarCraft Esports and other sporting events abroad, predictions and betting go hand-in-hand. With the help of community-sourced data of online and offline tournaments, I have created an app that enables users to generate predictions.
Language: Python(pandas, numpy, sklearn, xgboost) Framework: Dash
The Data
Aquisition and Formating:
The data source, Aligulac, has a simple method of extraction: PostgreSQL database dump. Restoring the dump file to a local database and exporting relevant tables to CSV files was made easy by the pgAdmin GUI.
Data Wrangling and Baseline Prediction:
Isolating a select feature set required a handful of useful Pandas operations. I filtered off games played before the LOTV expansion, created a new column to represent the winning player, and merged two tables to house the match and player rating information. A baseline classification prediction was made from the summary statistics.
The Model
Training and Inference:
The training and validation split was done at a 75/25 ratio. The data split was stratified across the target column to ensure equal distributions of classification values. To make inferences, an XGBClassier algorithm was chosen. Tuning the hyper-parameters of the model involved a round of randomized-search cross-validation using classification accuracy as the target evaluation metric. The trained model achieved an 82% accuracy score on the hold out data; a nearly 20% increase from the majority baseline.
Deployment
Dash and Heroku:
I trimmed the feature set by permutation importance, leaving the top two predictive features for the deployed model: Player A and B’s smoothed rating. These features had roughly 3X the predictive power as any other stand-alone data column. The Dash library makes bootstrapping the front-end simple. Getting under Heroku’s free-tier memory limit didn’t cause problems with the pickled model. Check the links below for the deployed app(give Heroku time to spin-up the server).
Restrospective
Hypothesis and TO-DO’s:
Firstly, I’d like to implement a pipeline of varying model types. I’m curious about the circumstances that lead to a lesser rated player winning, whether it’s more prevalent in a race v race scenario, and how the distribution looks across balance patches. Treating the data with a tighter time series window might also prove helpful. I can imagine a model trained with the most recent results having more predictive power. That being said, I’ve yet to test the model against new data.
Links: Web App — Github Repo