League of Legends Post-Draft Win Predictor
A credible win-probability signal, scoped to where it's real.
SITUATION
Champion select is widely believed to shape who wins a League of Legends match, but most public “draft predictor” tools claim to forecast the winner before the draft even happens — a much harder, noisier problem than the accuracy numbers they report can actually support.
TASK
Build a model that predicts the match winner immediately after champion select locks in — not before — using real ranked-match data, and be explicit about that framing so the accuracy number means what it claims to mean.
THE DATA
Champion-selection data from the Riot API captures rank, player, and pick information, but the signal that actually differentiates outcomes isn't the picks themselves — it's what differentiates two players on the same champion across rank tiers, which has to be engineered as a feature rather than read off the raw data.
ACTION
Used the Riot API to collect match data, engineered features from rank, player, and champion-selection differences, and trained a transformer-based model to predict the winner immediately after champion select, reaching 57.34% accuracy.
TRADE-OFFS
Scoped the prediction strictly to post-draft: 57.34% accuracy is a strong, credible result for a prediction made after champion select locks in, but the same number framed as a pre-draft claim would be much weaker and less credible. Keeping the framing honest mattered more than a punchier-sounding but overstated headline.
DESIGN DECISIONS
Post-draft, not pre-draft
The model always predicts after champion select locks in. Framing the same 57.34% accuracy as a pre-draft prediction would overstate what the underlying signal can actually support — the honest framing is a deliberate choice, not a limitation to hide.
One model output, two uses
The same win-probability output doubles as a matchmaking-quality signal — how balanced a given lobby is right after lock-in — rather than only being read as a pure win/loss prediction.
DEPLOYMENT
A data and modeling pipeline (Riot API → feature engineering → transformer training) rather than a hosted service — built to validate the modeling approach and framing, not as a deployed product.
RESULT
A post-draft win-probability signal accurate enough (57.34%) to be credible on its own terms, and honest about the boundary of what it can claim — a smaller, truer result instead of a bigger, unsupported one.