How our stats work

A harvester script pulls the top 200 players from 5 regional leaderboards (Global, US, GB, KR, BR) — about 984 unique players total. For each one, it fetches their last 25 battle logs from the Brawl Stars API and filters out anything that isn't real competitive play:

• ·Non-competitive modes (Showdown, 5v5 events, novelty modes)
• ·Friendly / practice games
• ·Battles with missing timestamps or no result field

Every valid battle writes a BattleRecord row for each of the 6 players in the game, with win/loss flipped per team. Then the harvester chains one hop deep — every unique player tag seen in those battles gets added to a secondary pool, and up to 400 of those get harvested too. So the dataset isn't just the top of the ladder, it's the competitive ecosystem around them.

Result: ~185,000+ BattleRecord rows, each tied to a specific brawler + map + result. Every row is raw and append-only — we never edit or delete this data.

The aggregator writes to two tables, and the split matters. An earlier version of this pipeline used a single table and was silently dropping ~18% of real battles because of a map-filter mismatch. Splitting fixes that.

Tracked competitive modes: Gem Grab, Brawl Ball, Bounty, Heist, Hot Zone, Knockout, Siege, Wipeout, Duels. Duels is 1v1, not 3v3, and included as a deliberate call — tank/sustain brawlers underperform there but it's still real competitive data.

Raw win rates lie when samples are small. A brawler with a 3-0 record has a 100% win rate but the signal is noise. Every displayed win rate gets pulled toward 50% using a Bayesian prior of 50 virtual 50/50 games:

shown_rate = (wins + 25) / (total + 50)

Effect at different sample sizes:

• ·3 wins / 3 games (100% raw) → 53.6% shown
• ·27 wins / 50 games (54% raw) → 52% shown
• ·270 wins / 500 games (54% raw) → 53.6% shown
• ·2700 wins / 5000 games (54% raw) → 53.9% shown

Big-sample brawlers barely move. Tiny-sample brawlers get pulled hard toward the middle so they can't top dashboards on noise alone. Pick rate isn't shrunk — it's a pure ratio and small samples don't produce misleading outliers there.

Tiers are the loudest badge on the site, so they need to mean something:

• ·S = win rate 54%+ AND total battles 1000+
• ·A = win rate 51%+
• ·B = win rate 48%+
• ·C = below 48%

The 1000-battle floor for S is the guard. At 1000 games, a 54% observed rate has a 95% confidence interval of about ±3pp — true rate is very likely 51%+, which is genuinely strong. Without the floor, a 500-battle brawler on a hot streak would get the same S badge as one with years of proven data. The rule references no brawler by name — everyone auto-promotes or auto-demotes as their data moves.

Per-map rankings don't sort by raw win rate — they sort by Wilson score lower bound at 95% confidence with a 50-game Bayesian prior. Wilson answers: "what's the lowest plausible true win rate given this sample?" It pulls uncertain samples down harder than certain ones, so 2-0 (100% raw) ranks below 44-33 (57% raw) because the first has too little data to trust. Same math Reddit uses to rank comments.

Three cards on every map detail page, picked live from the map's stats — not from a pre-computed field:

• ·Best first pick — top of the Wilson-sorted list, restricted to brawlers with a real sample
• ·Safest pick — win rate 51%+ AND pick rate 3%+. The consensus strong-and-popular choice
• ·High risk / high reward — win rate 53%+ AND pick rate under 3%. Niche picks punching above their weight

Any of the three can return empty, and the card simply doesn't render. We'd rather show two cards than fake a third.

Brawlers are typed: lane, tank, assassin, thrower, sniper. A counter matrix defines the competitive triangle (lane beats thrower, tank beats assassin, assassin beats lane, thrower beats tank, and so on).

When you select enemy brawlers, every available brawler scores:

• ·+2 for each enemy type it's strong against
• ·-1 for each enemy type it's weak against
• ·Small bonus from win rate as a tiebreaker

This is competitive theory, not empirical head-to-head data. BattleRecord only stores the harvested player's brawler per battle, not the full enemy roster, so we can't compute true matchup win rates yet. That's a planned upgrade once the schema captures all 6 participants per battle.

The draft simulator tracks ban/pick state for both teams, recommends highest-impact brawlers to ban (win rate × pick rate = danger score), runs the counter engine against enemy picks for top-3 suggestions with reasoning, and computes a live advantage score — your team's average win rate minus the enemy team's. Positive means you're favored.

• ·/debug/stats — per-brawler comparison of raw BattleRecord win rates vs what's stored. They share the same SQL source, so they should match within ±0.5pp. Any drift is a real bug. Also shows what numbers would look like under the old map-filtered aggregation path, with deltas — visual proof of the bias we fixed.
• ·/debug/maps — which map names from BattleRecord matched our Map table and which got dropped.
• ·The Brawl Stars API is public. Pull any player's battle log and check our data against it — they should match exactly.

Leaderboard players play the actual competitive meta. They draft intentionally, play optimal modes, and their results reflect real brawler strength — not casual chaos. Chain harvesting one hop deep pulls in the opponents and teammates those top players fought, so the sample isn't just the top of the ladder — it's the competitive ecosystem around them.