Prometheus
Prometheus is a database of 'sabermetric' like stats that evaluate League of Legends esports teams and players. The prometheus repository will contain a database that records these stats as well as a self-hosted website to browse stats across different teams, seasons, and regions.
An overview of prometheus' stats (described more in the metrics section). They come in two families. Season stats describe a team-season with hindsight and are judged on how stable they are and how well they match that season's results. Forecasts predict the next game from earlier games only and are judged by a backtest. No forecast uses a season stat as input.
Season stats:
- Global League Offensive Rankings Yield (GLORY): Prometheus' flagship metric. How well a team played: gold/objective stats weighted by their importance in that year's meta, with every game adjusted for the opponent's strength over the season.
- Record (sunset; still used inside GLORY): what a team achieved: its results over the season, adjusted for schedule, with fast wins counting for more.
- Luck (sunset): wins above what a team's play earned.
Forecasts:
-
Game-length Elo: a rating for every team in every region, built from player ratings (a team's Elo is the average of its five starters, so ratings follow players through roster moves). A new player counts 250 points lower in their first game, the discount fading with the log of games played until it is gone at 200, because teams with inexperienced players lose more than an average start predicts. Glicko-1-style player rating deviation (RD) scales each move by RD²/floor²: new/floor/reference 280/140/140, growth 10 per √day, information .125. RD grows while inactive and shrinks with games, replacing provisional and time-away multipliers; there is no volatility or g(RD). Non-major domestic updates use expected-score slope .0065; majors and international events keep ln(10)/400, K 20 everywhere. A player moving down from a major league to another league gains 25 points. Fast wins move it more: a fixed curve around 30 minutes makes a 25-minute win count as about 0.96 of a win and a 40-minute win as about 0.66. International results move shared league offsets; non-international cross-league games add team-only links (share .5), and compensated transfers add evidence (share .02 over 20 games). A separate forecast calibration adjusts international win probabilities using residual regional errors learned from completed earlier events; it does not change player ratings or offsets.
-
Player Elo: the player ratings behind team Elo, on their own register and one page per player (Elo after every game, career by team). Teammates share the team's result, with moves scaled by each player's RD, so it follows a player's teams through a career rather than splitting credit within a team.
-
FORGE (Form + Elo, formerly GlorELO+): the headline forecast of who wins the next game. Elo plus Form, a team's recent, opponent-adjusted stats weighted to predict the next game and compared with its own league.
-
Live in-game win chance: the feed's gold lead/share, dated Elo, raw objective counts and the change in gold lead over three estimated game minutes. The frozen objective+momentum model improves retrospective feed-map log loss by 0.0136 (95% series interval 0.0100–0.0170); its calibration error is about 1.8 points. The pre-map model is unchanged, and missing objective counters fall back to the earlier gold+Elo model. The clock is estimated and results do not establish a trading edge. Validation. A prepared draft-v2 API candidate records private parallel calls; displayed predictions stay on the current model.
Live-map trading (prepared, disabled): the selected 30-second trade-stability strategy combines the unshifted live model and last-trade price using earlier-only monthly fits. It requires 5 cents of edge after conservative fees and supports private shadow observation or capped-Kelly IOC execution through separate activation. Historical tape results are not executable-profit evidence. Implementation and rollout.
Sunset: Form (folded into FORGE; alone it predicts no better than Elo), GLORB (equal-weight baseline for GLORY; predicts no better than win % so far), GLORY+ (folded into GLORY), and the unadjusted GLORY.
CLI
The prometheus CLI provides users the ability to view past of current ratings for any of prometheus' metrics.
For example, the following command will show the top rankings of the GLORY metric for all major regions in a given year.
$ prometheus rankings glory --league MAJOR --year 2024
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ GLORY Rankings (MAJOR | [2024]) ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ teamname score year league │
│ 0 Gen.G 83.60 2024 LCK │
│ 1 G2 Esports 77.11 2024 LEC │
│ 2 Team Liquid 70.83 2024 LCS │
│ 3 T1 65.70 2024 LCK │
│ 4 Hanwha Life Esports 65.46 2024 LCK │
│ 5 Fnatic 60.97 2024 LEC │
│ 6 FlyQuest 60.41 2024 LCS │
│ 7 Team BDS 57.71 2024 LEC │
│ 8 Cloud9 53.86 2024 LCS │
│ 9 Dplus KIA 53.51 2024 LCK │
└────────────────────────────────────────────┘
Provide lists of arguments to filter results for all of those parameters. Years also take inclusive ranges, so --year 2021-2023 is the same as --year 2021 --year 2022 --year 2023.
The metric can be glory, record, luck, forge, form or the sunset glorb. Forecasts (forge, form) show current ratings by default; with --year, they show ratings at the end of each season:
以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗
