How we turn build choices
into MWPA scores.
The site is designed to answer one question cleanly: did this build choice actually help, or did it just show up more often in games that were already won? MWPA estimates the difference by comparing each option to an appropriate baseline — instead of raw global win rate.
The Mean Win Probability Added framework used throughout Goldiff is directly adapted from research by xPetu. The core thesis: raw win rates on items and runes are misleading because they conflate correlation with causation.
Without xPetu's work, most build sites would still be reporting that Mejai's Soulstealer has a 75% win rate and calling it a recommendation. MWPA cuts through that noise.
Start from the champion's own sample, so Ahri mid is not graded against random support or bot-lane fragments. Each role-champion pair has its own contextual baseline derived from matched game state at decision time.
Item paths are evaluated as first, second, and third checkpoints. Order matters because real purchase timing changes the signal — and because games that reach a third item have already self-selected.
Runes, shards, boots, and spell pairs add directional context on top of the item path instead of pretending one stat explains the entire match. The composite is a weighted sum across components, with confidence used to dampen tiny samples.
Item purchases are compared against alternatives bought by the same champion and role in similar slot, timing, gold, and win-probability states. Sparse choice sets fall back to standard MWPA instead of forcing a counterfactual read.
Counterfactual MWPA starts from peer-relative MWPA, blends available 2, 5, 10 minute and next-objective WPA windows, then penalizes wide intervals, weak peer overlap, low sample quality, and large initial-state selection bias.
Hidden from default views. Will surface only with explicit volatility opt-in.
Surfaced with a confidence band; treat directional reads as suggestive, not conclusive.
Treated as the primary read. Confidence is high enough that the contextual baseline is stable.