MMR Calculator

Appearance

ADR-0001: Three-tier inverse-variance MMR estimator ​

Status ​

Accepted

Context ​

Existing MMR tools (AussyELO, RankedKings, NoobHours) take a player's visible rank and multiply their win rate against a scalar constant. This produces a number that is entirely determined by visible rank — it cannot distinguish a player whose MMR matches their rank from one whose MMR is 400 points higher. The signal is weak and the output is fake-precise.

We need something that (a) uses the strongest available signal and (b) is honest about uncertainty.

Decision ​

We use a three-tier estimator:

Tier 1 — Win behavior (cheap, from league-v4 only): Computes the gap between estimated MMR and visible rank from the deviation of a player's win rate from 50%, scaled by sqrt(n) for sample confidence. Cost: 1 API call.

Tier 2 — Lobby composition (expensive but most informative): For each of the player's last 20 ranked games, we look up the current rank of all 9 other participants and compute a weighted lobby average. Opponents are weighted 0.6 vs teammates 0.4 (opponents are a tighter MMR signal). Recency-decays at exp(-days/14). Discards games with >2 unranked participants. Cost: up to ~180 API calls, heavily mitigated by the D1 ranked_entries cache.

Tier 3 — Inverse-variance combination: Both tiers produce (mu, sigma) estimates. We combine via w = 1/sigma^2, which automatically weights the more confident estimate. The resulting 90% CI is displayed to the user. sigma is floored at 40 MMR to reflect irreducible matchmaking noise.

Consequences ​

  • Better accuracy: Tier 2 is the only publicly available MMR signal that doesn't rely on the player's own rank. For steady-state accounts (20+ games), it dominates the estimate.
  • Higher API cost: Cold requests touch ~200 Riot API endpoints. With a dev key (100 req/2 min) this is throttle-limited. The D1 ranked_entries cache (24h TTL) means repeat checks of popular players are near-free.
  • Honest uncertainty: A new player or sparse account gets a wide CI rather than a fake-confident number. This is a feature, not a bug.
  • Future upgrade path: Tier 1 upgrades to true LP-delta inversion once 7+ days of daily snapshots exist for the account. The code path is ready; it just needs a flag check on snapshot count.