Trust through transparency

Where every number comes from.

We compute statistics from public esports data so coaches, analysts, and viewers can scout faster. This page documents our sources, sample-size policies, and refresh cadences. If a number on this site surprises you, this is where you check our work.

Data sources

Oracle’s Elixir

Historical professional drafts and player scoreboards underpin draft analysis, champion presence and player performance. Imported competition coverage differs by league and season.

Leaguepedia (Cargo)

Schedules, roster metadata and scoreboard builds. Player champion builds use unambiguous matches to the selected Oracle cohort; unmatched and ambiguous games are excluded.

Riot Data Dragon

Champion class tags, ratings, range and assets. Composition analysis pins a metadata version and discloses fallback snapshots.

Riot Match-V5 and live feeds

Additional match details and live features use their respective integrations. Live availability does not imply completed historical draft coverage.

How we read the data

1. 1

   Separate observations, reference data and predictions

   League, season, split, patch, team, side and dates select the observed cohort. Draft scores use a curated multi-league reference: LCK, LCKC, LPL, LEC, MSI, WLDs, FST, AC, CBLOL, CBLOLA, CD, LTA S, LTA N, LCP and TCL. They describe retrospective meta fit, not causal drafting skill or a win probability.

2. 2

   One scoring contract

   The composite weights champion strength 15%, lane matchups 20% and champion synergies 65%. Bayesian priors shrink samples toward 50, with strengths of 15, 30 and 20 respectively. Weighted wins and weighted games drive scoring; actual game counts describe samples. Unknown champions receive a neutral 50 prior. Ban strategy is a separate diagnostic, with opponent evidence also shrunk toward 50.

3. 3

   Explicit reference windows

   Patch reference includes the selected and previous available patch. Trailing windows end at the selected patch. Sparse pairs can expand through 5, 10 and 15 patches, then the year reference. Year champion, matchup and generic synergy references weight the selected season 1.5 and preceding season 1; role-specific duos use the selected season. Later results within a window may contribute, so historical scores remain retrospective. Score details name the fallback scope.

4. 4

   Sample counts and uncertainty

   Champion rates, relative team tiers and composition groups expose their observations. N/A means no observed sample. Samples below 20 are marked limited; this is a disclosure threshold, not proof that larger samples are reliable. 95% Wilson intervals are descriptive and ignore series dependence. Meta-fit ties are excluded from W/L rates and equal displayed rates share a tier.

5. 5

   Exploratory composition groups

   Groups use overlapping class tags, Riot ratings and range. They do not directly measure engage, scaling or counter strategy. Three fixed initializations compare candidate group counts, with minimum group sizes and a no-clear-grouping fallback. Both sides remain in the reference when filtering Team or Side. Large queries retain the latest 4,000 match contexts and disclose truncation. Group IDs are local to their cohort and metadata snapshot.

6. 6

   Player performance needs context

   A champion’s outcomes mix player skill, role, teammates, opponents and execution. Simulator player samples use actual counts within the selected year, league, patch and role. Player detail follows the profile filters. Frequent teammates describe shared games, not an estimated synergy effect. Archetype-adjusted KPI multipliers are heuristic context, not independently validated causal adjustments.

7. 7

   Predictive features must earn their claims

   Model candidates use a chronological holdout, a seven-day embargo, training-only feature vocabulary, calibration summaries and side/team-history baselines. Candidates are saved separately; old unvalidated artifacts are excluded from inference. A separate research backtest freezes year-reference inputs before each UTC match date. Neither evaluation establishes the predictive validity of retrospective patch-scoped tiers. Kill-prediction coverage labels describe available inputs, not calibrated certainty.

Refresh cadence