League Win Predictor

The question was whether a League of Legends game is effectively decided before it ends. I scraped the top 100 ranked players on the Korean ladder, pulled their match timelines through the Riot API, and captured each team's position at 14 minutes: gold, level, minions, jungle farm, kills, deaths, assists, turret plates, towers, dragons, heralds and wards.

I compared logistic regression, a decision tree and a random forest, each on the full 13-feature set and a reduced 10-feature set that drops plates and ward counts. Evaluation used a shared helper reporting precision, recall, a confusion matrix and ROC-AUC, so the models were judged the same way rather than on accuracy alone.

Logistic regression is the model worth showing interactively: inference is a dot product, so exporting the fitted coefficients lets the predictor run entirely in the browser with no backend. It reaches 79.5% accuracy and 0.886 ROC-AUC on games it has never seen.

The result that surprised me was how lopsided the coefficients are. Gold lead carries about seven times the weight of the next strongest feature — vision and plates barely register once gold is known, largely because gold already encodes them.