My previous post ended with “didn’t find that retail trader, probably never will.” This one is what happened when I actually ran the data: zero survivors. But the sentence more valuable than “didn’t find one” is this — the data does contain a statistically-significant trading edge, but it lives entirely in multi-coin systematic strategies, and not a single one of them is “a human making directional calls on a few instruments.” Real edge exists, just not in humans.
1. How I ran it
The previous post surfaced two datasets: Hyperliquid’s Info API (per-trade fills and funding rates all natively on-chain, no authentication needed), and CoinLobster’s 60 pre-filtered wallets (it has a “<30 trades/day” filter that conveniently excludes HFT bots). This time I took those 60 wallets as the address entry point and pulled full data for each from Hyperliquid: per-trade fills (with closedPnl, fee, crossed, coin, timestamp), the per-event funding cash flow, and the deposit/withdrawal ledger.
I wrote it as reusable Python: a client with pagination, throttling, and 429 backoff; pure functions for profiling and gate-checking; 9 unit tests pinning the PnL sign math and gate logic. The net profit I trust is the one I recompute myself: sum(closedPnl) - sum(fee) + sum(funding) — I don’t trust any leaderboard headline, because leaderboards exclude funding, and in perpetuals funding is the big one.
2. Seven gates, zero survivors
The previous post set seven gates, all required to count as “a trader worth copying”: over 2 years, net-profitable after funding, under 30 trades/day, under $1M account, not a market maker (maker<30%), concentrated on few coins (≤10), human-paced (trade interval >10 seconds). All 60 wallets run through it: zero pass all seven.
Two gates were the killers. Human pace — only 1/60 pass. Even the “low-frequency” wallets trade in sub-second bursts; that’s machine rhythm, not a person clicking. Coin concentration — only 10/60 pass; 50 wallets spread across 11 to 204 tokens, the fingerprint of multi-coin bots/arbitrage. Over 2 years — only 26/60, because Hyperliquid launched in 2023; the platform’s youth is itself a hard wall.
The sharpest detail: the three longest-lived wallets (3.0 to 3.4 years, the only ones that genuinely clear the 2-year wall) all fail on coin-spread (151 to 198 coins) or maker-ratio (0.38 to 0.58). The longest-lived are precisely the multi-coin market makers — confirming the previous post’s thesis that they all died on HFT/market-making fingerprints.
3. The funding recalculation’s surprise
49 of 60 are genuinely net-profitable once funding is added back, so CoinLobster’s ranking is mostly honest. But 11 wallets flip net-negative after funding: one is directionally −$3.08M and collected only +$40k in funding, landing at −$3.07M net; another is directionally −$5.7k but paid −$39k in funding, landing at −$54k. So the “leaderboards exclude funding” pitfall is real, and it bites the losers — people who were already losing directionally and then kept paying funding.
This bears directly on my own MacdCross backtest: funding cost must be actually computed, not conveniently ignored, or the backtested profit won’t match live trading.
4. The decomposition: directional alpha vs funding carry
I took the 21 “old + net-profitable” candidates (relaxing the machine-detection gates, just these two) and decomposed net PnL into two blocks: “directional (closedPnl minus fee)” and “funding carry.”
The result overturned my initial assumption. I had expected the winners to be funding harvesters, showing “profit” by collecting funding. The data says the opposite: 21/21 are directional-dominant — profit comes from closedPnl, not funding carry; 0/21 are funding-dominant. Funding is a minor cost for winners — a wallet net +$500k pays only −$3k to −$39k in funding.
5. The edge is real, but it’s all in the machines
I ran a sign-permutation test on each candidate’s per-close PnL: randomly flip the sign of each trade’s PnL 2,000 times and see whether the real mean is significantly more extreme than a “no-edge” random baseline. 16/21 are statistically significant (p<0.05) — not pure luck, not survivorship bias; there’s a repeatable positive expectancy. Win rate ~50%, payoff ratio 1.2 to 2.9, machine-grade consistency.
But — every one of those 16 with significant edge trades 84 to 204 tokens (essentially Hyperliquid’s entire universe), with ~50% win rate and machine-consistent payoffs. That’s cross-sectional stat-arb, momentum-across-universe, or market-making at machine scale. Three are maker-heavy (market making); thirteen are aggressive takers across the whole universe.
Zero of 21 are “concentrated on ≤10 coins + statistically significant” — the retail-directional-human archetype doesn’t appear, not once.
6. The methodological trap (the one thing to remember)
The per-trade significance test flags any positive-expectancy trading as “skill,” including a market maker capturing spread. So “16/21 have skill” read alone is misleading — it conflates market-making edge with directional alpha. You only get the truth when you cross with maker-ratio and coin-concentration: after the cross, all 16 significant wallets are multi-coin bots, and the count goes to zero.
This is a reminder for me too: don’t mistake a machine’s systematic edge for “a retail trader’s skill” and try to copy it — that’s copying the wrong object. Edge has to be honestly attributed.
7. Pattern extraction: what the edge actually IS
The previous sections left an open question: of those 16 multi-coin machines, is the edge cross-sectional momentum, funding arbitrage, or mean reversion? I took each wallet’s trade structure apart one more layer and checked all three.
First, rule out funding arbitrage. 16 of 16 are net funding PAYERS (negative carry); zero systematically hold the funding-receiving side. Funding is a cost for them, not the edge. That kills the “they look profitable by harvesting funding” hypothesis stone dead.
Next, rule out “sub-second market making” and “simultaneous cross-sectional arbitrage.” The median hold time is 35.6 hours (hours to days), not the second-level flipping of a market maker. Only 1 of 16 is a majority-maker (maker 0.78; two more hybrids at 0.3–0.5 are aggressive-leaning-with-some-making); the rest are aggressive takers (maker median 12%). The decisive number is cross-coin simultaneity: only 3.4% — they enter one coin at a time, sequentially, not by rebalancing the whole universe at once. So not the “portfolio-rebalancing bot” archetype.
The edge is: multi-coin, aggressive-taker, trend-following/momentum, run at swing cadence. 14/16 have a momentum-shaped PnL distribution (11 are textbook trend-follow: payoff 1.5–2.9, winrate <55%, right-skewed — frequent small losses, occasional outsized wins; another 3 are “both-high” — high winrate AND high payoff but extremely right-skewed, dragged up by a few outlier wins). Only 1 is mean-reversion-shaped (winrate 57%, payoff 1.22). They scan the whole coin universe (median 127 coins), take taker positions one coin at a time, hold ~1–2 days, and the PnL distribution is the textbook trend-follow fingerprint.
The direct measurement (upgrading momentum from fingerprint to measurement). I reconstructed each coin’s on-chain price series from the px field of all 60 wallets’ fills (no CEX, no candles — Hyperliquid’s candles REST endpoint 422s on every format; the only official path is the AWS bulk export), then for each wallet’s every OPEN fill computed the coin’s pre-entry return over four windows (1h/4h/24h/48h) and measured the correlation between entry direction and pre-entry return: + = momentum (long after up, short after down), − = mean-reversion (long after down, short after up). 92% of opens have a same-coin prior fill within 24h to anchor the reference price — near-total universe coverage.
The result is more honest than the fingerprint — and partly overturns my shape-based inference. The entry signal is SPLIT, not uniformly momentum: at each wallet’s strongest window, about 9/16 align momentum (positive corr) and 7/16 align mean-reversion (negative corr), with a median correlation of only +0.05 (the 48h window, the strongest) — the effect is real (p<0.05) but small. The sharpest detail: several “momentum-shaped PnL” wallets (right-skewed, payoff>1.3) actually ENTER mean-reversion (buy after the coin went down) — e.g. 0xb99b strongly negative −0.46, 0x0533 −0.31; and the one market-maker 0xf8e6 is the strongest mean-revert enterer (−0.56, consistent with market making — providing liquidity, getting filled passively).
This reconciles shape and measurement. A right-skewed PnL (frequent small losses, occasional big wins) does NOT imply momentum entry — it comes mainly from exit management (let winners run, cut losers fast), which BOTH momentum-entry and mean-reversion-entry traders can produce. So my reading “edge = momentum” from the PnL shape was an over-read; the shape does not determine the entry signal. The unifying edge is NOT the entry signal (it’s split) but asymmetric exits: whether they enter momentum or mean-reversion, the winners all use “let winners run, cut losers” to make the PnL right-skewed.
The hard warning for MacdCross. MacdCross is a MACD crossover (momentum entry), but the data shows “momentum entry” is NOT these winners’ universal edge — exit management is. So my strategy’s edge depends more on exit/stop discipline than on the entry signal itself; whether you enter momentum or mean-reversion matters less than “cut fast, let run long.”
Consistent with my whole project’s direction: don’t believe get-rich myths, get the costs and funding right first, build a falsifiable system, and let it die in front of you on historical data.