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):
- Global League Offensive Rankings Yield (GLORY): Prometheus' flagship metric. Weights gold/objectives by their importance in the meta, and calculates the best teams at securing those advantages across all regions.
- Global League Offensive Rankings Baseline (GLORB): Baseline for GLORY. Weights all objective/gold equally.
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.
$ prometheus rankings glory --year 2021 --year 2022 --league LPL --league LCK
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ GLORY Rankings (LPL, LCK | [2021, 2022]) ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ teamname score year league │
│ 0 Gen.G 77.18 2022 LCK │
│ 1 FunPlus Phoenix 76.00 2021 LPL │
│ 2 Royal Never Give Up 74.74 2021 LPL │
│ 3 Top Esports 74.40 2021 LPL │
│ 4 T1 72.67 2022 LCK │
│ 5 Dplus KIA 72.38 2021 LCK │
│ 6 EDward Gaming 64.63 2021 LPL │
│ 7 Invictus Gaming 63.38 2021 LPL │
│ 8 JD Gaming 62.46 2021 LPL │
│ 9 Suning 60.10 2021 LPL │
└────────────────────────────────────────────┘
- One advanced metric (GLORY) that evaluates team performance
- One advanced metric that evaluates individual player performance
- SQL database that stores
match, player, andstattables - Self-hosted webpage for exploring stat rankings
- CLI for dynamically viewing metrics
- Metrics span 5+ years of data over all major regions
These metrics evaluate the strength of a team. Their main purpose is to compare the strength of a given team compared to teams across regions.
These metrics are heavily dependent on the strength of opposing teams and the meta. For example, a team that plays during the 2018 skirmish-heavy meta will tend to have higher scores than the control-meta of 2015, regardless of the truth strengths of their team.
To account for this, Prometheus calculates stats based on a per-season basis. Each metric is weighted appropriately for a single season. Thus, teams are scored based on how well they accomplish the goals of that specific "meta".
Prometheus also includes z-scores to compare the strength of a team relative to other teams in its region. We can calculate which teams were most "dominant" in specific eras.
GLORY is prometheus' flagship stat for team performance. At a high level, it is a weighted sum of basic per-game stats listed below.
- GPM = team_gold / game_length_minutes
- GDPM = (team_gold - opponent_gold) / game_length_minutes
- Turrets_per_10 = turrets_taken / (game_length_minutes/10)
- Baron_per_10 = barons_taken / (game_length_minutes/10)
- Dragon_per_10 = dragons_taken / (game_length_minutes/10)
- Objective_conversion_rate = objectives_taken / contested_objectives
We also include slightly more advanced metrics in this calculation:
- Kill_conversion = fraction of team kills that directly preceded (within X seconds) an objective take
In GLORY, weights are calculated using logistic regression with the winning team as a prediction target.
GLORB is very similar to GLORY, except all categories are weighted equally. It provides a baseline to compare GLORY against and is much easier to implement.
TBD
Determining the weights for WOBA involves creating a table where each match has two rows: one for Blue side and one for Red side. The features in a row are the stats described above.
The match data will be stored in an sqlite3 database. Initially, the only table in the database will be a match table that stores the weights for LORB/GLORY.
When a user requests the metrics for a certain team, region, or split, we will query the database to grab the relevant matches for that metric request. Then, we will fit a model using scikit-learn. The weights for those models will be cached or stored on the disk, depending on the time it fit the model.
The Prometheus API will return metrics and rankings for use in the CLI and website.
Unlike traditional sports, most modern League regions don't follow a traditional round-robin based format for their regular season. For example, here is the LTA North Split 1 season:
As a team advances farther in the bracket, they face more difficult opponents, which results in worse metrics. Stronger teams that advance farther are "punished" by the metrics for winning their games and moving through the bracket.
One solution to this problem is to factor in opponent team strength when calculating metrics. This can be simply by using opponent win percentage to adjust scores, or calculating a team's elo.
Weighting by elo will also help the model learn weights that reflect the tendencies of top regions as opposed, which is the ultimate goal of these metrics. Take the following contrived example.
T1 (LCK) is undefeated during the spring split. Every game, they sack all dragons and only focus on taking voidgrubs and heralds. C9 (LCS) is also undefeated. Every game, they sack voidgrubs/heralds and only focus on taking dragons.
Because T1 is playing in a more difficult region, intuitively their games (where top-side is prioritized much more heavily) should 'count' more in the models. Thus, our model should favor teams that are great at taking voidgrubs/heralds over their counterparts that focus on dragons.
If we weight matches by the opponent's elo, then matches against stronger opponents will count more in the rankings, solving this problem. The difficulty with this is implementing an elo system that works across regions and across the 10+ league seasons.