2026-10-05

Trench lessons — 2026-10-05

Key takeaways

How profitable memecoin traders actually make money
Finding and vetting copy-worthy wallets, and the traps
Why copy trading fails, and fixes that work
Exit rules for memecoins
Rug, bundle and sniper signals before entry
MEV, Jito tips, priority fees, landing

How profitable memecoin traders actually make money

2026 (Apr data, Dune live to Jul): Share of pump.fun+PumpSwap wallets with positive realized PnL went 30% (Jun 2025 low) → 50% (Jan 2026) → 73% (Apr 2026) → ~80% (Jun 2026). Most winners made $1–500. Only 5.4% made >$1K. Realized only: bagholders who never sold are excluded, so this overstates real profitability.
Nov 2025–May 2026: A 6-month look at 6.8–7M wallets: about 44–49% profitable, median realized PnL ≈ $0.02–0.06, mean ≈ $900–1,160. The edge sits in a thin tail, and the typical wallet earns nothing.
Paper, data Jan 2024–Jan 2026 (pub. Sep 2026): "Meme Coin Factories" (CMU/EPFL), all 15.2M pump.fun coins: ≥17% of trades are wash trades (atomic buy+sell in one tx). Coins with 500+ wash txs graduate ~10x more often. Top 1% of creator clusters (funder-linked, 3 hops) made 58.6% of all coins. About 8k coordinated "many wallets → one dumper" sells per 1% sample. ≥10% of coins are copycats, and the original graduates at 9.2% vs 0.86% for copies. Only 1.02% of coins graduate.
USENIX Sec 2026: Across 34,988 memecoins on 4 chains, 82.8% of high-return tokens showed artificial growth (wash trading or LP-based price inflation) before pump-and-dumps or rugs.
WWW 2026 (UCL/NTU): Copy trading is "the dominant entry strategy" in memecoins. On 6,000 coins, the best model picked smart wallets averaging +14%/coin, but the copier earned only +3% after frictions. Statistic-only models (XGBoost etc.) found profitable wallets yet still gave negative copier returns. Wallet-level signals mattered most (ablation −11% AUC), then coin-level, then timing. Trader profitability is predictable, but the edge is small after copying.
Mar 2026 KOL cohort: Among 332 tracked KOL wallets, net flow was roughly zero (rotation). The highest-volume wallets were not the most profitable. Some low-trade-count wallets had the best profit per trade (e.g., 448 trades for +560 SOL).

Finding and vetting copy-worthy wallets, and the traps

WWW 2026: Wallet gates that predicted profitable next trades: t-stat of per-coin returns > 1.645, mean return > 0, return SD < 1, enough trades (top 75% of sample), positive returns in each horizon (last 1–5, 6–10, 11–15 coins), recent activity, and older wallet age.
Jun 2026 (GMGN): GMGN's own vetting list: compare 7D vs 30D, check avg hold time (seconds-long round trips mean a bot or farmer), and look at phishing flags: bought honeypots, "Not Purchased" (transferred-in) share, sell volume > buy volume, and buy-sell within 5s flips. Trace funding. GMGN PnL omits priority fees, tips and failed-tx gas.
May 2026: Bot-wallet tells: failed-tx rate (bots fail far more often; >15–20% is a yellow flag, >40% red), mechanical timing, single-token concentration >35–40% of volume, wallet age vs sophistication, recurring same-window co-entries, and win rate >70–75% over 50+ trades (multi-wallet routing or transfer-inflated PnL). Cielo, Birdeye, Solscan and GMGN don't auto-filter bots.
2026: Wallet types to avoid: farmer wallets (big buy into a thin pool, copiers pump it, they dump), whales whose size moves the price, and HFT wallets.
2026: Leaderboard PnL is easy to manufacture (showcase winners, park losers in side wallets, wash). Some "top traders" are deployers trading their own launches.

Why copy trading fails, and fixes that work

2026: Detection delay on GMGN copy is typically reported at 2–5s. High slippage settings invite sandwiches.
Jun 2026: GMGN: pick one exit owner: follow-sell OR your own TP/SL (setting TP/SL disables follow-sell). Use fixed size or a max cap, plus market-cap, liquidity, token-age and max-copy filters.
Mar & Aug 2026: Convergence ≠ conviction. High-KOL-count tokens often had negative net flow (crowd in and out). In Aug, across 533 KOL wallets, the median 2nd KOL followed the 1st within 9 seconds, at a median first-buy market cap of ~$4.7K. Many "clusters" are fast followers or linked wallets, not independent conviction.
Jun–Jul 2026: 1,012 persistent sniper cohorts (2–12 wallets, 2,965 addresses) co-fire as early buyers across launches. A corrected re-analysis shows only a small effect on later buyers, so the cohorts exist but their price impact is modest. Treat co-firing wallets as one entity for clustering.

Exit rules for memecoins

2026: Dev wallet "Transferred Out" on a fresh token is the single strongest rug signal. Dev selling ≥50% is a fail.
WWW 2026: "Dump duration" (peak until liquidity falls to 10% of peak) is the key risk clock. Bundled coins dump faster.

Rug, bundle and sniper signals before entry

WWW 2026: Bundle test: any non-creator buy in the creation block (Alg. 1). Sniper test: non-creator buys in the first 5 blocks (snipers are everywhere and mostly just shift profit). Bump-bot score: flips ÷ net position change ≥ 50. "Gradual bundles" show mechanically regular candles. Bundles were in ~25% of coins and linked to shorter dumps.
Feb 2026: MemeTrans dataset (pump.fun, Dec 2024–Mar 2025): bundled accounts held 36.5% of supply. After grouping bundles, the top-10 concentration of high-risk tokens rises by a median of 24 points. Cluster bundles by shared funder, tx ID, or Jito bundle ID.
2026: Bundle-scanner practice: ignore 2-wallet same-slot pairs at low %. Worry about 3+ wallets in the same slot still holding real supply (e.g., 24 wallets holding 76% means it's going to 0). CEX-funded bundles can evade funder links.
2026: On pump.fun, mint revoked and LP locked are the default (99.9% and 72.9% of 18.9k flagged rugs had them). "The rug is in the wallets," not the contract.
Sep 2026 paper: Coins with no "pump" suffix were likely created outside the UI by scripts (16% baseline, 26.5% among copycats). Copycat name/symbol/image is a negative signal unless the coin is the original.

MEV, Jito tips, priority fees, landing

2026: Jito BAM (TEE block assembly) reduces sandwich exposure but doesn't remove it.
2026 docs: Jito bundles: tip ≥1,000 lamports, ideally inside the same tx. Bundles are for atomic multi-tx work, not for a single swap.

Sources

Evidence grades: A = large dataset, B = official docs, C = vendor/thread/small sample. Research notes for education only. Not financial advice.