SCPredict: StarCraft II Replay win rate analysis and visualization

SCPredict is an online analysis tool for StarCraft II replay files.

Model and Training Data Used

SCPredict is trained on the official posted WCS/ESL Masters replay files from 2020 to May 2024. A total of 4445 replays were used to train the models.

Zephyrus Replay Parser was used for 1v1 replay parsing. One model for each in-game matchup was trained seperately, as the features used in the model is unique to matchups. Logistic regression models were fitted using scikit-learn's Logisticregression() module with L2(Lasso) penalties.

Model Performance

The classification report for the Protoss vs. Terran model:

	precision	recall	f1-score	support
0.0	0.68	0.87	0.77	10601
1.0	0.75	0.49	0.60	8412
accuracy			0.70	19013
macro avg	0.72	0.68	0.68	19013
weighted avg	0.71	0.70	0.69	19013

Note that each game start with the highest entropy possible, which gets reduced as in-game events unfold.