---
name: high-prob-patterns
description: Ben's weekly high-probability pattern hunt (from 4 Oct 2026) - the method, the patterns that survived out-of-sample tests, and ones that collapsed. Read before proposing any "high win rate" setup.
sources: [cowork]
aliases: [high probability patterns, weekly pattern, win rate patterns, deep oversold, dip in uptrend]
---
- [stated] 4 Oct 2026: Ben wants ONE extremely high-probability pattern at a time (rare is fine), found fresh each week, rotating money from one to the next; filters can include chain, coin age, market cap, MA state, social data.
- Method: one row per coin per day on 4h candles (GMGN), 1,481 coins (937 main set RH/BSC/Sol + 544 never-seen Solana), 24h volume >= $10k, meme-like volatility only. Rules of up to 3 conditions searched on the older half of the main set, then checked on the newer half, on half the new coins, and finally on the other half of the new coins (untouched). Exits on 4h closes, max 14 days, 2% fees.
- Lesson: discovery win rates of 60-70% collapsed to 13-36% on new data for most rules. Win rate also depends on exit shape: with +10% target / -20% stop, random entry already wins 55%; with +15/-30 it wins 50%. Judge lift over random, and average trade after fees.
- C "deep oversold" (best): price 31%+ under 4h MA200, no more than 46% under daily MA200 (so coin is 200+ days old: BSC/Solana), 4h RSI(14) <= 33.5. Exit +10% / -20% / 14 days: 80% win (129 trades, 96 coins), untouched holdout 76%, main 78%, new 83%, every month 71-86% except May (n=6); avg trade +5.8% after fees vs random 55% win. About 1.5 signals per week. TESTING.
- AB "dip in uptrend": 4h MA200 up 10%+ over 7 days, price down 22%+ over 7 days and down 4.4%+ last 24h, 20%+ under 4h MA100, still 45%+ above 30-day low. Exit +15% / -30% / 14 days: 70% win (200 trades, 186 coins), main 69%, new 71%, avg +7.6% after fees vs random 50%. Works on RH too (75%). About 2.3 per week. TESTING.
- No pattern reached 85%+ on untouched data. Social (X) data was not in this run (no history for thousands of coins).
- Scripts: feat.py (features + labels), mine.py / mine2.py (search with holdouts), check.py; data F2.pkl, F3.pkl in session; candles Mac ~/ma/gm2.tgz and gm3b.tgz.
## Money printer search (2026-10-04)
- [stated] asked for a rare, high upside "money printer" pattern (once a week or every two weeks is fine)
- Price/MA/age/chain rule search aimed at +50% and +100% targets (returns capped at target, mcap excluded): nothing survived the final untouched holdout. Best rule (age <=181d, within 18% of ATH, daily RSI <=40) went +17% early, then +0.6% coin weighted on holdout.
- Listing news after the public post (@NewListingsFeed, 218 events, 104 tickers): no edge. Binance Alpha trade -10.8%, Upbit -11.3%, Bithumb -9.9%, Robinhood spot -7.7%. Robinhood asset add: +228% in the 24h BEFORE the post, then -31% by day 3. Samples tiny (n 1 to 9).
- Conclusion: big upside needs info earlier than price and public posts (pre-listing wallet flow, insider buys, X mention speed-up). Still best tested edges: C deep oversold and AB dip in uptrend.