Our methodology

The Player DNA is our unified profiling system. It extracts 35 behavioral features from your last 30 ranked games (grouped into 6 chromosomes: Combat, Macro, Mechanics, Mental, Team, Tempo) and classifies you into one of 13 archetypes, all hand-validated (Sniper, Berserker, Architect, Ghost, Showman, Guardian, Coinflip, Tyrant, Rat, Predator, Phoenix, Wall, Patient).

Each feature is z-scored (deviation from the global mean in σ), then we identify your 3 signatures (recurring strengths, top-3 positive σ) and your block (recurring weakness, most negative σ). The archetype is assigned by rule-based scoring: score = mean(z[positive keys]) − mean(z[negative keys]) (a mean, not a sum — so archetypes with 3 or 5 keys stay comparable). Plan A (HDBSCAN centroids on 100k EUW vectors) sits in the Sprint 4 backlog.

How a metric becomes a score

Every raw metric (CS/min, vision, damage…) is converted into a 0-100 score through linear interpolation against a rank-calibrated benchmark, then adjusted by a role multiplier. This is the machinery that powers the detailed match grade (next section) and the Player DNA z-scoring.

The score is then clamped between 0 and 100.

Benchmarks by rank

The bounds are calibrated on real data extracted on March 23, 2026 from our database (830k+ ranked games, EUW1). Manual refresh on major patches — a weekly scheduler is on the roadmap:

Role multipliers

Benchmarks are adjusted per role to compare a support and an ADC fairly:

Every match gets an overall grade (S, A, B, C, D) based on 6 weighted categories. The grade is computed relative to your rank, not in absolute terms: a B in Diamond is far better than an A in Bronze.

The 6 categories

* The damage share is role-adjusted: the reference 8-35% range is multiplied by the role played (UTILITY × 0.45 → 3.6-15.75% range, BOTTOM × 1.15 → 9.2-40.25%, etc.). Before this fix, a tank support at 10% damage share was capped around 20/100 on this component (25% of Teamfight) no matter what — now a support at 12% scores the same as an ADC at 28%.

⚠️ These ranges are reachable targets per role, not an additive breakdown: the per-role medians don't sum to 100%. A team's damage share sums to 100% by construction (5 players split 100% of the damage), but each role has a different envelope for what it's expected to weigh individually.

Category weights per role

The 6 weights above are the laner weights (Top / Jungle / Mid / ADC). A support structurally has neither income (gold/CS) nor objective last-hits: judging them on that tanks their grade no matter what. The UTILITY role therefore redistributes those weights toward what a support actually controls.

How to read a grade (post-match)

Since 2026-05-20, every sub-grade (Laning / Combat / Survival / Objectives / Vision / Farm) in the match summary is clickable. Opening the detail shows you:

• Every sub-metric with your value vs the benchmark for your rank + role and a ▲ / ─ / ▼ status.
• The sub-metric's normalized 0-100 score (colored mini-bar) and its weight % in the category grade.
• A ● marker on the benchmarks adjusted to your role (see the multiplier table above), with a “Calibrated for <role>” banner on top.
• An “X weighs the most in this grade” bubble pointing at the sub-metric responsible when the grade drags down — handy to know what to work on first.

Grade thresholds

Percentile by rank (interpolated)

The raw 0-100 score is converted into a percentile against players of the same rank via stepwise linear interpolation. That avoids discrete jumps — a 79 score is no longer abruptly at 78% and an 80 at 90%: the function lerps between anchors for a continuous progression.

Impact map (impact within the game)

The Impact map (impactScore, 0-100) measures your real weight in that specific game, not against your rank. It is role-aware:

For UTILITY, objective participation (a share of team damage — a support does 5-15% of it no matter what) is replaced by vision and presence, which actually measure their presence on the map:

MVP / LVP

The MVP badge goes to the player with the best globalScore (0-100) among the 10 players. The formula is role-aware:

For UTILITY, the formula drops CS/Gold (irrelevant for the role) and redistributes toward vision and impact:

The KDA contribution is capped at 8: without that cap, a Perfect KDA (e.g. 6/0/28 → KDA 34) saturated the score at 100 and killed MVP discrimination between Perfect-KDA players. impactScore itself is role-aware (see Impact map above, volume-weighted KP included).

Carry — every role in its own way

• Damage Carry — laners who carry through kills: kill share ≥ 50% of the team, ≥ 5 kills, in a win.
• Engage Carry — supports who carry through engage/peel/vision: UTILITY + win + KP ≥ 65% + ≤ 3 deaths + ≥ 1.5 vision/min. Before 2026-05-09, the generic "Carry" badge structurally excluded supports.

“God of X” game badges

Every game also hands out “best of the game” badges on a given metric. Since 2026-05-30, the “God of X” badges require absolute floors on top of being the best — so a 1/3/1 in a stomp doesn't grab “Support God” just for being the only support on their team:

• Support God — best supportScore of the game, and ≥ 10 assists and ≥ 25 vision.
• Vision God — best vision score, and ≥ 25 total and ≥ 1.4/min (a long game with mediocre vision no longer cuts it).
• CS God — best CS/min among laners, ≥ 7 (supports excluded).

Overall grades give an average, but a splitpushing Top and a carry ADC don't share the same responsibilities. The Role Signature is a composite score specific to each role, built on 5 KPIs with different weights per role.

The 5 signatures

Every death is tagged and scored from Riot timeline events (positions, nearby allies/enemies, wards, objectives, HP before death, gold diff). The system outputs a 0-100 score and a category (good / acceptable / bad / terrible).

The 13 detected tags

Quality score

Base at 50. Each tag applies a bonus/penalty:

Contextual modifiers

The 14-day progression plan is generated directly from your scoring diagnosis, with no AI call. Before (until April 12, 2026), a Claude call invented the objectives and thresholds — expensive, slow (~15s), and sometimes inconsistent. Now it's generated server-side in pure TypeScript: instant, free, and deterministic.

How it's generated

Catalog of 12 objective templates

Each template defines an extractable metric from the match data, a target, and 3 hand-written actionable tips:

Objective auto-validation

Every time you visit your Progression page, the system re-fetches your latest matches and automatically checks each objective via extractMetric():

No need to click to validate — as soon as you play a game that meets the condition (e.g. 8 CS/min), it's automatically counted. You watch your progress update in real time.

Skill tree

The Progression tab also shows a micro-skill tree (recall timing, survival, objective focus, wave management, vision…). Since June 2026, each grade is computed server-side from your real timelines: your ~15 latest ranked games run through the same analyzers as the rest of the site (death analysis, wave management, macro, pings) and the aggregate grades each skill on what actually happened in game.

Transparency rule: if a skill is not measurable on your recent games (missing timeline, missing data), it's shown as “unavailable” — never padded with a neutral grade that would fake it. The tree's overall grade only averages the skills that were actually evaluated.

With a sample of ~1,000 games per champion, raw winrate is too volatile. A champion at 60% winrate over 80 games isn't necessarily better than one at 52% over 3,000 games.

The small-sample problem

With little data, rarely played champions show extreme winrates (40% or 60%) through pure statistical variance. We combine two complementary estimators: Bayesian smoothing to estimate the true WR, and the Wilson score interval (95% lower bound) to penalize small-sample uncertainty.

1. Bayesian winrate formula

2. Wilson score interval (95% lower bound)

The Bayesian estimates a central point but says nothing about uncertainty. Two champions can share the same bayesianWR with radically different confidence intervals. The Wilson score gives the lower bound of the 95% CI — a champion with 50 games (σ ≈ 7%) mechanically ends up below a champion with 1000 games (σ ≈ 1.6%).

with z = 1.96 (95% CI), p̂ = observed winrate as a fraction, n = games.

3. Final score

Mix of 70% Bayesian (global stability) + 30% Wilson (penalizes uncertainty) + popularity bonus:

The bonus rewards popular champions (meta-validated by volume) with no positive cap. The penalty has been floored at −3 points since 2026-05-20: the Bayesian regression already handles small-sample reliability — beyond that, we would double-punish viable pocket picks (a champion at 200 games in a pool averaging 800 took −6, about a full tier on top of the regression).

The Bayesian/Wilson mix stabilizes tiers from patch to patch: “lucky outlier” champions climb less.

Tier assignment

Champions are sorted by descending score, then split by percentile:

The minimum threshold is adaptive: between 10 and 50 games to be included, based on the pool's average volume. On a freshly released patch (little data) the threshold drops toward 10 so the list doesn't empty out, then climbs back toward 50 as collection catches up — Bayesian/Wilson pulls small samples toward the 50% prior anyway.

The theoretical rank estimates the rank you should be at by crossing your performance with that of other players at your current rank. Since 2026-05-12, it is powered by the Player DNA: we no longer aggregate per-rank thresholds over your averages, we directly read the z-scores the ML service computes across 35 features (see /api/ml/dna).

Algorithm

Weighting (sum = 1.0)

Confidence score

Below 20 ranked games in the DNA window, no estimate is produced (per-bucket z-scores too noisy). sampleFactor (0 to 1) measures how solid the archetype assignment is — it's not a simple games ratio. Confidence is bounded between 35% (thin sample, soft signal) and 100% (full sample, clear signal > 1.5σ).

The Tilt Detector analyzes your recent session (consecutive games < 90 min apart) and detects 7 tilt signals.

Score calculation

31 badges across 9 categories, each with 4 progressive rarity levels. Thresholds are fixed and absolute (not rank-relative), but three badges are adjusted by main role so every role can unlock them at its natural level: damage_dealer and gold_income are normalized by role multiplier (a support at 330 DPM = an ADC at 660 DPM = Bronze), and cs_machine is locked for UTILITY mains (CS is irrelevant — they already have ward_master, control_ward, ward_killer).

Each badge has 4 tiers: Bronze → Silver → Gold → Legendary. Progress toward the next tier is shown as a percentage.

The system automatically detects problem players (AFKs, inters, smurfs) on each team to gauge the luck factor of your games.

How we separate your real contribution from the team's result. A win proves nothing on its own — a player can carry with a 20/0 in a loss, or get carried at 1/8/15 in a stomp. Here's how LegendsTracker attributes performance to the individual rather than the collective scoreboard.