S16 V4.1: broader practice coverage, unchanged parameters

中文 · English

V4.1 inherits all V4 models, features, patch responses and small corrections. Only professional practice inputs change: additional OBGG-listed Worlds-player accounts across servers, with OBGG ownership accepted by default. Original V4 forecasts and inputs are preserved.

Complete V4 method, equations, historical errors and limits. The archived39/95 coverage figure describes the old snapshot; current coverage appears below.

Combined: 3821 player-game observations, 2841 eligible main-role records and 980 excluded autofill records; 60/95 starters have eligible data. 1460 additions comprise 1060 main-role and 400 autofill records. 45 China accounts are excluded by request; 49 non-China accounts remain unresolved.

Different pros in one match are distinct player-practice observations. Each player/server/match appears once. A hero enters the practice pool only when actual role matches the official role. OBGG participant rows run Top/Jungle/Mid/Bot/Support; identity prefixes do not establish played roles.

weight = 2^(-age_days/10)

share(player,hero) = (weighted_games + 20 × professional_pick/2) / (weighted_total + 20)

practice_candidate = 2 × mean_over_19_starters_in_role(share)

practice_adjustment = clip(0.05 × (candidate - professional_pick), -0.02, 0.02)

Accounts pool per person; missing players retain professional priors. KR Master+ remains at most5%; numeric/mechanism patch response remains1.25, system1.0. Only remaining patch shocks apply, preserving effective game volume. Novel roles still need3 games/1 pro or existing KR gates and cap at0.5 percentage points. Role picks sum to2 each, bans to10, hero pick+ban≤1. Aggregate mean BP positions weight role picks; ranked games provide no draft sequence.

S15 coefficients were not retrained, and V4 auxiliary priors were not tuned. Revisions reflect input coverage, not reduced error. No actual Worlds error, verified accuracy gain or new S15 backtest error is available.

Reproduce: run forecast_s16_worlds_v4_1.py, then render_s16_forecast_v4_1.py. Audit with audit_obgg_account_reference.py --skip-raw (full replay needs the source archive) and audit_s16_forecast_v4_1.py. Frozen repository data supports offline forecasting. Extract source responses into the clone using stored paths, then omit --skip-raw. Do not backfill past windows with present practice.

Coverage · Source replay · Model replay · Manifest

Largest pick revisions from coverage (percentage points, not errors)