Method

How the forecast works

We first estimate each team’s current strength, then run the official format one million times to convert those strengths into a realistic path to the title.

Team strength: Elo ratings

Every team has a rating. Winning against a stronger team gains more, losing to a weaker team drops more, and each new season nudges ratings back toward the mean to reflect roster changes, patch impact, and form swings. The gap between two ratings is the starting point for the series-win forecast.

01Each series updates the rating
02Ratings reset toward the mean at the start of a new season
03The rating gap drives the probability

Why B5 was chosen

We compared six candidate models: coin flip, historical win rate, basic Elo, tuned Elo, inactivity-decay Elo, and season-reset Elo. They were all tested on the same real matches using out-of-time validation, and B5 produced the lowest error, so it became the main model.

Prediction error

B0Coin flip0.693
↓ + team history
B1Historical win rate0.646
↓ + Elo dynamics
B2Basic Elo0.631
↓ + temporal tuning
B3Tuned Elo0.629
↓ + inactivity regression
B4Inactivity-decay Elo0.628
↓ + season regression
B5Season-reset Elo0.626

Out-of-time validation

We only use the past to predict the future: each round trains on matches that happened before a given point, then predicts the next section of the season, avoiding any look-ahead. This process covered 11 seasons and 1,330 series.












22S1	22S2	23S1	23S2	24S1	24S2	24S3	25S1	25S2	25S3	26S1	26S2

TrainPredict

Tournament simulation

We take the 12 team ratings and replay the official format from Stage 1 through the Breakthrough, knockout, and final, repeating the full tournament one million times. The share of simulations that reach each stage becomes the path and title probability.

1,000,000

full tournament simulations

Stage 1
→
Breakthrough
→
Knockout
→
Grand Final
→
Champion

All forecasts were fixed before the first match on 2 October 2026; results do not rewrite them.