GPO VALUES trade calculator

The full pipeline, step by step

Below is everything the system actually does, top to bottom. If you want the version that fits in a text message, see the paragraph above.

1 · Scrape

A Discord scraper pulls messages from seven trading channels (trade-1, trade-2, booster, premium, certified, plus event channels) and writes rows of (timestamp, message_id, author_id, content, reactions, reply_to) to gzipped CSVs. The message_id column lets us dedupe exactly across scrapes so adding new data never double-counts. Author IDs let us throttle bumpers + wash-traders.

2 · Parse (Rust)

A custom Rust parser turns messages into structured trades at ~700K messages/sec (roughly 1.4s per 1M messages). Each message gets tokenized through a 3-tier fallback: Discord custom-emoji ID → emoji name → text-alias lookup. A 1,100+ entry alias table maps community shorthands (pcc, pchakram, chakram, chak, sos, lih, etc.) to canonical item names.

A compound resolver with ~300 rules assembles multi-emoji item names: Exalted + Wings → "Exalted Cupid Queen's Wings"; Legendary + Fish_Bait → "Legendary Fish Bait"; Kraken + Blade + Azure → "Kraken Blade (Azure)"; Max + Legendary → max-stack leg fish bait.

Two dedupe layers kill bump-spam:

• Exact dedupe by message_id — duplicates across scrapes removed.
• 60-second bump filter + per-user-per-day cap — a bumper reposting the same trade every 5 minutes contributes at most once per day.

"Offer me / LF X / How many for Y" open-ended posts are dropped (no ratio = no value signal). ":Peli:" (in-game currency) is filtered out as noise.

3 · Scam filter

Pure 1:1 same-quantity trades are dropped from the ratio graph entirely. Cross-tier 1:1s like 1 PCC FOR 1 SGC are almost always scam lowballs; filtering them protects the Dijkstra math below. Cross-item ratio info still comes through multi-quantity trades + bundles.

4 · Solver (Rust)

Anchor: Mythical Fruit Chest = 10,000. Every other value is derived as a multiple of this anchor. The Rust solver runs in ~9 seconds on 1.8M trades. Five phases:

• Phase 1 — ratio chain: Dijkstra finds the max-bottleneck (highest-evidence) path from the anchor to every item. Each edge is a weighted-median ratio across all trades between those two items. Scam 1:1s dropped, pure-ratio trades kept.
• Phase 2 — bundle propagation: multi-item trades like PCC for PFL + PWE give "1 PCC = sum of those two." For each item we take the weighted-median of implied values across all bundles where it appears alone on one side.
• Phase 2b — iterative re-valuation: since item values depend on each other, we iterate with EMA damping (35% new + 65% old) and a 30% per-round step cap. Converges in 10 rounds.
• Phase 2c — balanced windows (per-item): each item uses up to 500 most-recent trades (250 offer-side + 250 want-side) as its own evidence pool. Prevents high-volume items from drowning out niche items. Uses the 40th percentile for items with ≥40 weight of bundle evidence (guards against a handful of "flex offer" bundles inflating the median). Aggregate clamp: items can't move more than 1.6× from their pre-phase value in one solver run.
• Phase 2d — soft tier-ordering penalty: if a lower-tier item exceeds any strictly-higher tier's p75 by >20%, pull it down 10% (single pass). NEW and VERY_UNSTABLE tiers exempt.

5 · Community anchors

A small number of items have community-validated values applied as high-weight pins (data/suggested_values.json). These override the chain math when the weight is ≥500 — used for items where public Discord data systematically diverges from known community consensus (fruits, mid-tier items). Low-weight community suggestions blend proportionally.

Anyone can propose one via /suggest in the Discord server (with 2+ proof screenshots). Auto-approves at 3+ ✅ reactions with 1.5× yes/no ratio.

6 · Time-decay + trust weighting

Every trade carries two weights:

• Long window: 30-day half-life — a trade from 30 days ago is worth half as much as today's.
• Short window: 12-hour half-life — picks up intraday market shifts.

Per item the solver picks whichever window is more trustworthy right now (window_active on each card). Authors flagged by our wash-detector get their trades downweighted (0.5×).

7 · Post-process refinements

• 1:1 alignment — items stuck at a bad chain estimate get pulled halfway toward a dominant 1:1 partner (≥50 trades).
• Demand adjustment — items the community wants more than they offer get a boost of 3–20%; over-offered items get a matching discount.
• +Adds learning — the value of "+Adds" modifiers is learned per trade-scale tier from mod-bearing trades. Lets us re-include previously-dropped mod trades with a learned correction.
• Wiki injection — every item is cross-referenced with the GPO Fandom wiki for rarity, drop rate, and item images. Powers the rarity-frame glow + detail panel.

8 · Confidence bands

Every item is count-driven, not human-adjusted:

• High confidence — 1000+ trades. Range is tight (±10%).
• Medium confidence — 200–999 trades. Range is wider (±15–20%).
• Low confidence / ⚠ LOW DATA — <200 trades. Amber badge; range is wide (±30–40%). Treat the number as a ballpark.

We never override the band to make an item look more or less certain — the math decides.

9 · Daily history backfill

Starting from the first trade date, we replay the solver at 23:59 UTC each day using only trades available as of that day. This gives every item a 60-day price time series we can chart — without having needed to run the solver daily ourselves. Used by the Changes tab and detail-panel trend arrows.

10 · Discord bot

GPO VALUES exposes all this in chat — natural-language parsing ("pcc for pfl"), 30+ slash commands, personal inventory tracker, price alerts, trade challenges, interactive help hub. Channel-gated per guild; details at gpovalues.com/bot.

What's intentionally NOT here

• No editor's opinion — if the data says X, it's X.
• No paywalls, no premium tier — every value is free forever.
• No fake confidence — thin-data items are flagged prominently.

Caveats

• These are trade offerings, not closed deals. The math weights what the community is asking, not what actually clears hands.
• Exclusives (Marine Cap, Valkyries) trade rarely — holders don't post, so offers under-sample them vs. community consensus. The LOW DATA badge marks this.
• Brand-new items (first-seen <30 days ago) swing hard for the first two weeks — use the /history chart to see the trajectory.
• "+Adds" at the prestige tier can mean another full prestige worth of fillers. The learned adds model handles this but individual trades still carry that ambiguity.

Open source + feedback

Report incorrect values with /feedback in the Discord server. Propose corrections with /suggest (community vote approves). API access at api/items.json (free, rate-limited). See the FAQ for common questions, confidence explained for how to read LOW DATA, or bot command reference.