S16 Worlds champion selection & BP forecast
Frozen pre-event forecast · S15 shared parameters + S16 historical features · Worlds outcomes are pending. Whole-event scope includes Play-In, Swiss and Knockout.
- Cutoff: 2026-10-09T19:20:41Z
- Patch: 26.20 / Data Dragon 16.20.1
- 2,556 games; 1,003 series; 40 events; 173 champions × 5 roles.
Model and transferred parameters
S15 V1 weighted historical priors plus a V2 ridge residual layer. Preserve 22 features, six frequency regressions, five position regressions and their coefficients, means and standard deviations. Ridge=0.03, residual blend=0.75, position blend=0.5. No S16 or Worlds parameter refit; in-event strengths 20/30 are inactive.
Weights include time decay, patch distance, importance, pre-game mean Elo, event balancing and first-game weighting. LCS inherits the LTA N prior of 1450. Clubs initialize by their home league; cross-region events do not initialize all clubs equally. Elo uses preceding results only, K=20.
Event importance: MSI 1.8; EWC 1.4; First Stand 1.3; qualifiers/Road to MSI/regional finals 1.45; playoffs/grand finals 1.35; KeSPA/DCGI 0.75; promotion 0.4; others 1. League priors: LCK1600, LPL1580, LEC1500, LCS1450, LCP/CBLOL1430; outside challengers1350. See forecast_summary.json.
Numeric patch effects use favorable-direction mean log(new/old) × ability importance × role relevance. Costs/cooldowns invert the sign; attack speed uses total multipliers and mitigation uses remaining damage. Mechanism effects, item exposure and representative combat stats are declared assumptions. Only main Summoners Rift changes are encoded; Classic/ARAM/Arena are excluded.
Projection enforces two expected selections per role, ten picks, ten active bans and champion pick+ban≤1. Position moments satisfy E[step]=E[rank]+6+4P(rank≥7). Roles come from played lineups; bans have no final lane. Ordered draft slots determine selection priority independently of map side.
Temporal checks
13 windows / 210 games.
| Metric | V1 | V2 |
|---|---|---|
| Pick-rate RMSE / pp | 5.334 | 5.381 |
| Presence RMSE / pp | 9.216 | 8.615 |
| Champion × role RMSE / pp | 2.434 | 2.460 |
| Pick-weighted rank MAE | 1.467 | 1.462 |
Groups require ≥30 games. Earlier 65% complete series provide preceding information, later series provide ≥10 test games. All other events are also cut off chronologically. Matching historical Data Dragon supplies attack range, rather than 16.20 values. New-patch shocks are zero for within-observed-patch continuation checks. The fixed universe is 173 states with cutoff-dependent new-hero masks. Metrics average windows equally; this protocol is not directly comparable to the S15 validation split.
Uncertainty and scope
All 173 live champions are assumed eligible; the event-specific restricted list has not been obtained. LOS/FURIA have sparse history. No completed 26.19/26.20 samples are available; patch effects are extrapolated.
- Event-specific disabled roster not obtained; all 173 champions assumed eligible.
- CBLOL-only games excluded by requested collection scope: LOS/FURIA team preferences are estimated from sparse cross-region observations.
- No completed 26.19/26.20 games in this dataset; target-patch shocks are extrapolations.
- Future series lengths, stage participation and within-series unavailable pools are not explicitly simulated; this is the same marginal whole-event model as S15. First-game weighting moderates observed Fearless depth.
- Independent game-level source checks cover only part of the archive; no claim of every draft step verified against video.
- Historical-series bootstrap measures sampling variation, not all model, patch or event-format uncertainty.
80 bootstrap draws resample entire historical series with fixed shared coefficients. Sensitivity scenarios remove patch effects, vary numeric effects ±25%, or remove mechanism/system priors. They are not separately trained models.
No fixed Worlds game count is assumed: outputs are whole-event per-game marginal rates. The four-team double-elimination Play-In changes stage mix; future lengths add risk. This preserves the S15 marginal model and does not treat an unknown schedule as fact. Later actual outcomes and evaluations must be saved separately, preserving this forecast.