Jimmy Tran

A five-model PyTorch ensemble that predicts how many turrets fall in a professional League of Legends game, and the betting record I built around it: 1,395 graded positions over 27 months, +5.74% on stake. The inputs weren't in any public API, so I collected them by hand through Google Sheets and Apps Script, and where an automated source existed I checked it against 1,793 games I had counted myself: 85.3% exact agreement, and a statistically significant undercount in the feed (p < 0.001).

The interesting part is that the model is barely predictive. Its out-of-sample R² is −0.002, slightly worse than guessing the average every time, and the returns come from pricing rather than forecasting. I also found a data leak in my own validation: random splits were letting later games inform earlier predictions, which inflated my headline number from +1.44% to +9.59%. I published the smaller one.

Reads professional broadcasts to capture what no API publishes: summoner spells and their cooldowns, items (about 85% agreement with Riot's data), kills, deaths and assists (97%), level (95%), creep score (90%), and every champion's position on the minimap, tracked every half-second. It learned a digit and icon set per broadcast, and each one is graded against Riot's feed.

Widening the match windows and carrying last-known positions raised kill placement on the map from 66% to 91%, checked against the broadcast camera itself: 99% of placed kills (104 of 105) fall inside the camera's view at the moment of the kill. It feeds the platform live, with roster checks so another stream's game can never leak in; the maps, summoner spells and quest items in the esports panel come from its logs.