S15 Worlds Draft Model V2

English dashboard · Chinese dashboard · Transfer parameters · Detailed training errors

Scope and definitions

All primary outputs cover the entire 84-game tournament: 4 Play-in, 52 Swiss, and 28 Knockout games. There are 171 champions and 855 champion-role states. Pick rate is picks / games. Draft presence is (picks + active bans) / games. Fearless unavailability is not counted as an active ban. Bans have no observed final role. Mean position is conditional on the champion being picked: global pick rank 1–10 or full draft step 1–20. Zero-pick actual means are null. Overall pick rate sums the five roles; overall mean position is weighted by predicted role pick rates. S16 has not been forecast.

Model

The model combines a weighted historical-statistics prior, shared ridge-regression residual models, and empirical-Bayes-like in-event shrinkage updates. The prior weights recency, patch proximity, event importance, pre-game average Elo, event sample size, and first-game status. It includes qualified-team preferences, ban pressure, and role-specific numerical/mechanism/system patch features.

The residual layer has 22 features, six frequency outputs (five role pick rates plus champion ban rate) and five role-position outputs: 253 regularized coefficients. Features include recent trends, event-team coverage, early draft priority, role entropy, nonlinear patch bases, range, and complementary-role co-selection. Champion identity is not a feature. Co-selection is an association, not a proven causal synergy. Outputs satisfy ten expected picks, two picks per role, ten bans and at most one pick-or-ban presence per champion.

Training uses 39 historical event-by-patch windows, containing 1,771 target games. Of 34 chronological validation folds, the earlier 27 select hyperparameters and the last seven are reserved from that selection. Each fold uses features only from before its start and fitted target windows ending before its start. Event fields are assumed known at start. Earlier historical balance changes before 25.19 are not fully encoded.

An in-event update combines prior and observations: (20 × prior rate + n × observed rate) / (20 + n). Mean position uses a corresponding pick-count-weighted prior with 30 equivalent prior games. After 84 observed games, roughly 80.8% of the frequency estimate comes from actual results, so the low full-posterior error is largely expected from shrinkage toward observed outcomes.

Whole-tournament results

The after-56 all-tournament score includes the 56 already observed games. It is not an independent forecasting score. Full-posterior results use all actual outcomes and must not be presented as pre-event accuracy.

Reused Knockout evaluation

Knockout outcomes were already inspected in V1, so this is a reused retrospective test, not a new blind holdout. New hyperparameters were selected only on historical events. Pure pre-event pick-rate accuracy is not uniformly better than V1. Seven Knockout series provide limited evidence; the paired bootstrap interval for mean pick-rank improvement crosses zero. See series bootstrap.

Champions and offline assets

English champion names and titles use Riot Data Dragon 15.20.1. The Chinese page uses Tencent Mainland China official names and titles. Both pages search English names, Mainland names, titles and the retained aliases. All 171 portraits are local PNG files. Role colors and data are identical in both languages. Click a bar, scatter point or table row for the same role details.

All champions in English · Asset manifest

Verification and reuse

The pre-Worlds archive has 2,392 verified games. The 84 Worlds games were cross-checked across OE, GoL and Leaguepedia. Rendering logic, segment sums, search, selection, dialogs, and exports are checked locally; the browser tool blocks file URLs, so no live browser layout inspection was performed. Exported SVG/PNG previews are standalone charts, not browser screenshots.

Shared coefficients and update rules are transfer priors only. Rebuild S16 historical features, available champions, patch changes, participants, and draft rules. Do not copy S15 champion posterior rates directly into S16.