We aggregate the live positioning and order flow of 37 consistently profitable Hyperliquid traders into a concentration-capped, signed per-coin consensus, the Mood Ring. A maker-first copybot mirrors that consensus into non-custodial follower accounts through approved API agent wallets, every order carries a 0.05% builder fee, and a snapshot bot keeps a strict 48-hour, time-indexed SQLite history of the entire pool. Resting exchange stops, a trailing stop, a re-entry block and layered watchdogs bound downside. This paper details the math, execution, fee economics, risk controls, competitive landscape and roadmap.
Short version: /whitepaper · LaTeX source: whitepaper-extended.tex
37 HL traders ──WS──▶ listener.js ──▶ Postgres (events, positions, accounts)
│ │
profiles.js (6h) ◀──┘ ▼
sentiment.js: 60% positions + 40% flow,
40% concentration cap ─▶ s_c ∈ [-100,100]
│
snapshotter (5 min, pre-cycle) ◀────────────────────┤
SQLite 48h strict window ▼
cycle-db.js (~5 min): w_c = s_c/100, |s_c|≤15 → 0
│
┌──────────────────────────────┼─────────────────────┐
▼ ▼ ▼
stop sync (8% SL, ALO at best bid/ask ─60s─▶ IOC 5 bps builder fee
trail 4/4, block) full closes → IOC → builder wallet
Conventional copy trading binds a follower to one leader. The follower inherits that leader's edge but also every idiosyncratic risk: a bad week, a blown-up position, tilt, a changed strategy, or simply the leader leaving the platform. Performance-ranked leaderboards also suffer survivorship bias: the top of the table on any given month is partly luck, and followers arrive just as mean reversion begins.
The Intelligence Consensus Architecture inverts this. Instead of copying a person, it copies a measurement: the net, dollar-weighted positioning and order flow of a pool of 37 independently profitable traders. Individual errors partially cancel; shared conviction survives. This is the same principle behind ensemble methods in machine learning and behind crowdsourced-signal funds such as Numerai (Section 10): many weakly-correlated, individually-skilled forecasters combined under a concentration cap produce a more stable signal than any single member.
Three design rules follow: (1) no single trader may dominate a coin's score (a 40% concentration cap); (2) positions and fresh flow both matter, so the signal reacts to new conviction without whipsawing on noise; (3) weak consensus is treated as no consensus (a flat band).
The listener (listener.js) subscribes over WebSocket to 37 Hyperliquid accounts and writes their fills, positions and account values to Postgres tables hlw_events, hlw_positions and hlw_accounts. The set is curated, not auto-ranked: accounts were hand-selected for sustained, multi-month profitability, active (day or swing) trading, and diversified style. It includes seven ZEC whales — large, consistently profitable Zcash traders who count fully in the signal.
profiles.js (re-run by the listener every 6 hours) builds hlw_profiles: a fictional nickname and style label per account (e.g. "Chip Whale — stock-perp position trader", "HYPE Monk — HYPE-only active trader", "ZEC Whale Delta — consistent ZEC long swing"), top-4 coins by volume, monthly PnL and a consistency measure. Rankings on /daytraders sort by these statistics; the ranking is descriptive and does not change signal weights.
Aggregate PnL of the pool:
Styles span intraday scalpers on majors, stock-perp traders on the xyz builder-deployed markets (NVDA, GOOGL, MU, INTC, SOXL), commodity traders (Brent, gold), pair traders and multi-day position holders. This heterogeneity is deliberate: it lowers correlation between members.
The signal is computed in api/_lib/sentiment.js over a fixed coin universe: BTC, ETH, SOL, HYPE, ZEC, NEAR and the xyz markets GOOGL, SOXL, NVDA, BRENTOIL, GOLD, MU, INTC.
Inputs. For trader \(i\) and coin \(c\): net signed position notional \(p_{i,c}=\sum \text{ntl}\cdot\operatorname{sign}(\text{szi})\), and 24-hour signed flow \(f_{i,c}=\sum_{\text{buys}}\text{ntl}-\sum_{\text{sells}}\text{ntl}\).
Concentration-capped score. For any map \(x_i\) with gross \(G=\sum_i|x_i|\), each contribution is clipped at \(\lambda G\) with \(\lambda=0.4\) (when more than one trader is present):
\[S(x)=100\cdot\frac{\sum_i \operatorname{sign}(x_i)\min(|x_i|,\lambda G)}{\sum_i \min(|x_i|,\lambda G)}\]
A trader holding at least 40% of gross is flagged as dominant in the output.
Blend. Position and flow scores combine 60/40, renormalised when one side is empty, then rounded and clamped:
\[s_c=\operatorname{clamp}_{[-100,100]}\Big(\frac{0.6\,S(p_{\cdot,c})+0.4\,S(f_{\cdot,c})}{0.6\,\mathbb 1_{P}+0.4\,\mathbb 1_{F}}\Big)\]
Labels: \(\le-60\) strong bearish, \(\le-20\) bearish, \(<20\) neutral, \(<60\) bullish, else strong bullish. Scores are persisted to hlw_sentiment at most every 3 minutes, giving a 24-hour history per coin.
Target weights. The copybot maps scores to signed weights with a flat band:
\[w_c=\begin{cases}s_c/100 & |s_c|>15\\0 & |s_c|\le 15\end{cases},\qquad N_c=w_c\cdot E\cdot L\]
where \(E\) is follower equity, \(L\) the follower's leverage, and \(N_c\) the target signed notional (subject to per-coin caps and the $25 minimum order).
Followers. Each follower (row in cb_followers) names a master account, an approved Hyperliquid API agent wallet, leverage and per-coin settings. Live followers today: Blue Bear (0x95Df…, 5x) and Smurf (3x). Agent wallets can place and cancel orders but cannot withdraw; their keys are stored encrypted (envelope-encrypted with a server KEK) and decrypted only in-memory for one cycle, then dropped.
Auto-live. With COPYBOT_AUTO_LIVE=1 the bot sends orders whenever a follower is active, the agent is approved and unexpired, and the builder fee is approved (maxBuilderFee >= 50 tenths of a bp). Otherwise it records a dry-run plan with a reason ("agent not approved/expired", "builder fee not approved").
Cycle flow. run-copybot.sh runs cycle-db.js roughly every 5 minutes:
Maker-first execution. Opens, adds and reduces for maker-enabled masters (COPYBOT_MAKER_MASTERS) are posted as ALO (add-liquidity-only) limits at best bid (buys) or best ask (sells). Unfilled orders are cancelled after ~60 s and the remainder is sent as an IOC market order with a slippage limit; leftovers under $1 are dropped. Full closes always go straight to IOC market for certainty. Today about 9 of 12 opens/adds filled as maker; the remainder fell back to market as designed.
Hyperliquid's base tier charges takers more than makers; at this tier makers receive no rebate — they simply pay less. All-in cost including the 5 bps builder fee was measured at:
On $10,000 notional: $6.00 maker vs $8.50 taker, saving $2.50. Worked examples at a 75% maker fill rate (blended ≈ 6.6 bps):
Of every fill, 5 bps goes to the builder; the builder share is constant regardless of maker or taker.
Every order the copybot sends carries builder: {b: 0x95e1642B6C06E06E6ad9412A90e908Dcb09f6c1a, f: 50}, i.e. 50 tenths of a basis point = 0.05%. Hyperliquid requires each user to sign an ApproveBuilderFee action authorising a maximum fee for that builder; perps are capped by the protocol at 0.1%. The bot reads maxBuilderFee every cycle and refuses to trade live if approval is below 50. Rewards accrue on Hyperliquid to the builder address and are claimed on-chain. Accrued to date: about $59 total. The model aligns incentives: revenue scales with volume routed, not with taking a share of follower profits.
The snapshotter (/workspace/hl-ws/snapshotter) runs every 5 minutes, scheduled just before the copybot cycle. Each run rotates current → previous, then captures the whole pool into a SQLite store (store.js, WAL mode):
snapshots(ts, meta, blob) — full compressed JSON per capturepositions, sentiment (score, weight), rankings, bot_state (equity, positions, stops), builder (rewards)fills — deduplicated by fill id, kept 72 h so the oldest snapshot's 24 h window stays completeRetention is a strict 48-hour rolling window enforced by prune() on every write. At full window (~576 snapshots) the store is expected at roughly 60–80 MB. A watchdog restarts the process on stall.
Query paths (snap.js): at <time> (nearest snapshot ≤ time), range <from> <to> (filter by address or coin), diff <t1> <t2> (position, signal and stop changes between cycles), and blob <ts> (full raw state). This gives an auditable, time-travel record of exactly what the crowd held when each trade was made.
COPYBOT_SL_PCT), 10% slippage cap, builder fee attached; replaced if size changes, cancelled on flatten. Because it rests on the exchange, it fires even if the bot is down.TRAIL_ARM_PCT), the stop moves to breakeven +0.1%, then trails 4% behind the best price; it only tightens. State lives in .sl-state.json.COPYBOT_CYCLE_MAX_MS) logs the stuck step and exits; a 10-minute heartbeat watchdog restarts the loop.Stops, trailing and re-entry block currently apply to Blue Bear only; Smurf runs without them by choice.
Hyperliquid analytics. HyperIndex, OAK Research, SmartPerps, WhaleRoom, HyperStats and CoinGlass provide wallet tracking, whale alerts and leaderboards. They show what smart money holds but mostly stop at information; execution is left to the user.
Auto-traders. HiveMind (open source, github.com/fzheng/hivemind) and TradeJS consensus explore multi-wallet consensus execution; HyperDash offers strategic copytrading of up to 3 wallets.
Single-leader CEX copy. Bitget (~1.1M followers), Bybit (800k+), Binance and OKX copy-trading, and eToro (35M+ users) dominate retail copy trading. They are custodial and typically take a 10–30% profit share for leaders.
Crowd-consensus evidence. Numerai aggregates thousands of anonymous models into one meta-model. Reported returns: 2020 +6.15%, 2021 +12.16%, 2022 +18.75%, 2023 −17.38%, 2024 +25.45% (Sharpe ~2.75); AUM grew from ~$150M (2022) to ~$1.9B regulatory AUM (Sept 2026). Figures are partly third-party reported and not independently verified. The 2023 drawdown is a reminder that consensus reduces, not removes, risk.
Advantages. Non-custodial (agent wallets cannot withdraw); consensus removes single-leader risk; a flat builder fee replaces opaque profit share; every order is on-chain and auditable; the snapshot store provides a 48 h audit trail of the crowd state behind each trade.
Risks.
Not investment advice. No keys, secrets or database URLs are included in this document.