How the four exit presets would have done
Every one of the 22,022 recorded launches bought at the scanner's entry price and sold by each Ghost Hive exit preset, after a 3% round-trip cost. These are estimates; the method is below the chart.
Running total, $100 a trade, all launches in time order
The horizontal axis counts launches, not days. A few early test launches date from April.
Model picks vs everything
How sure is this?
Low, mid and high are the volatility assumptions (10th, 50th and 75th percentile of real memecoin 5-minute moves). A result that changes sign across them is not a result.
Method
Who holds a token says a lot about whether it rugs
For each launch, Ghost Hive matched the top holders to the tokens those wallets held at least 24 hours earlier, so no later data is used. Holder history is the new signal here. Entry features such as market cap also predict rugs out of sample, and holder history adds to them. It is a risk check, not a buy signal.
Rug rate by holders' past rug rate
Bands are the holders' past rug rate; bars are how often those launches rugged. Source: ANALYSIS_v3.
Out-of-sample test
Rug model trained on launches before 2026-09-24, tested on launches after.
| Model inputs | AUC | Safest fifth | Riskiest fifth |
|---|---|---|---|
| Entry features only | 0.763 | 15.5% | 77.5% |
| + wallet history and holder concentration | 0.803 | 13.6% | 86.2% |
| Wallet history only | 0.737 | 19.0% | 76.6% |
Entry features: market cap, 5-minute volume, age, wave label, scanner score. AUC 0.5 is a coin flip, 1.0 is perfect. The rug label is Ghost Hive's own.
Other holder signals
Launches whose largest holder held over 40% rugged 43-44% of the time; the 20-40% band rugged 66%. There is no steady link. Holder history is the signal that held up in testing.
Check a wallet against the serial-rug list
wallets were top holders of 3 or more tokens with a known outcome, and every one of those tokens rugged. The busiest one held 52 tokens. Addresses were not checked against known pool or bonding-curve programs. Being on this list is a record of past holdings, not proof of intent.
All wallets with 3+ scored tokens
Share of their tokens that rugged.
Some very busy wallets may be infrastructure (pools, bonding curves), not people. No filter is applied: the list includes every wallet that meets the rule.
What the data says
Ghost Hive ran from April to October 2026 and is switched off. This page replays what it recorded. Nothing here trades, connects to a wallet or calls an AI model.
No filter or take-profit rule in ANALYSIS_v3 made money after costs
375 entry filters were picked on the first 60% of launches and tested on the last 40%. The filter ANALYSIS_v3 picked as best (market cap $250K to $1M, wave W3, take profit at +50% within 1 hour) averaged -4.6% per trade out of sample when recomputed from the shipped data (423 trades, 95% CI -13.0% to +3.7%). That is not distinguishable from zero, and it is not a profit. All 60 take-profit rules tested had a negative average. The only positive estimate on this site comes from a different kind of rule, the TRAIL10 trailing-stop exit preset, and it is only an estimate: +11.7% a trade at typical volatility, -2.9% at high volatility. See Exit presets.
85.5% of launch entries rugged or died
Ghost Hive's own summaries only counted tokens that still had a price. 98% of launches with no 1-hour price are rugged or dead tokens. Booked at -90% instead of dropped, holding for 1 hour returned -30.3% on average and 4 hours -56.6% (after 3% costs, missing prices booked at -90%, moves capped near +1,000%).
A higher score meant more rugs
The learned score's top tenth rugged 71% of the time, the bottom tenth 25%. Model picks did worse than unpicked launches once rugs were counted.
Smart-wallet alerts arrived after the move
Matched to the entry just before the alert, 1-hour returns looked like +26% to +54%. Matched to the next entry after it, when you could actually buy, they were -17% to -21% per trade (347 alerts).
Holder history ranks rug risk
Rug rate climbs from 27% to 86% with the holders' past rug rate. Adding wallet history and holder concentration lifted the rug model from AUC 0.763 to 0.803 on launches it had never seen (ANALYSIS_v3). The useful output is the rug-risk check and the 114,982-wallet table. It is not a buy signal.
Exit presets are estimates
Build info
Where the numbers come from
- ghosthive_launches_labeled.parquet: 22,022 launches with 5m to 24h moves, peak, drawdown, rug/dead labels, scanner scores and holder history.
- solana_wallet_rug_reputation.parquet: 114,982 top-holder wallets with tokens held and tokens rugged.
- Exit rules copied from the Ghost Hive paper engine source.
- Full write-up: ANALYSIS_v3 (2026-10-26).
Users
Recent access log
Newest first, up to 200 lines.