How Our Dota 2 Model Reads Drafts and Predicts Winners
By Esports Oracle · June 5, 2026
Dota 2 punishes overconfident models. Ours is built to be honest about that — Elo-led, draft-aware, and calibrated to make fewer big calls but get them right.
What carries the prediction
Leak-free factor testing was clear: Elo difference is the strongest signal (weight 0.16), followed by overall win-rate and ranking gaps. Head-to-head was over-weighted in earlier versions and has been trimmed; team strength does the heavy lifting.
The draft layer
When you supply the ten heroes, a draft-analysis factor (0.14) reads hero familiarity, meta alignment and signature picks for each side. It refines the call — but it never overrides a clear rating gap, which matches how pro Dota actually plays out.
Calibrated on purpose
We tuned the model's confidence down from an earlier version that said 85% when reality was 77%. The result is well-calibrated probabilities: on its high-confidence picks it still reaches 70%+, but it makes fewer of them — because in the most upset-prone game we cover, knowing what you don't know is the edge.