Strategy graveyard
Every approach we tested between September and October 2026, how we tested it, and why it did not make the cut. 21 entries.
| Strategy | How we tested it | Result | Verdict | Why |
|---|---|---|---|---|
| All-in market orders on dashboard signals | Paper, live prices, Sep 6–11 | −44.2%, 10% wins, 12 s median hold | Failed | Round trip cost 0.46%, edge 0.02–0.08% |
| 1–10 minute signal windows | 1.91M rows, 5 days | 58–63% direction right, +0.04–0.09% moves | Failed | Moves smaller than costs |
| Unconfirmed emergency exits | Paper ledger, 190 exits | −₩4.42M, 86.5% of all losses | Failed | A code path sold on raw alarms |
| Chasing coins already surging | 63-hour replay | −62% | Failed | Buys late, pays costs, catches reversals |
| Riding the day's early leader | Replay, 32 trades | +8.1% total | Inconclusive | Too few trades to trust; dropped |
| 34 classic indicator rules | Upbit history, hourly bars | 34 of 34 lost | Failed | Edge below costs |
| Oversold mean reversion | Upbit history | High win rate, negative expectancy | Failed | Small wins, large losses |
| Cross-sectional altcoin momentum | 252 variants | 6% of variants positive | Failed | Pump chasing |
| Time-of-day and weekday patterns | Upbit history | +0.14% a day before costs | Failed | Below costs |
| Zero-fee maker trading on BTC/USDT markets | Volume check | ₩2.3B and ₩0.6B a day | Failed | Market too thin |
| Copy trading and AI bots | Public record search | No verified track records found | Failed | No evidence of an edge |
| Local language-model trade picking | 87 events, 6,536 cases, 1-hour horizon | Its favourite 10% averaged −0.79% after costs vs. −0.44% for taking every trade | Failed | Ranked trades no better than chance (rank correlation 0.01) |
| Daily channel ensemble on top-volume coins | Replay | −62% | Failed | Top-volume lists chase pumps |
| Machine-learning predictions | Walk-forward | Lost in 2022, 2025, 2026 | Failed | Weak predictive power (IC 0.03–0.13) |
| Trend ensemble on 8 coins chosen in 2026 | Replay 2018–2026 | +75%/yr, fell to +41–46% point-in-time | Inconclusive | Survivorship bias |
| Kimchi premium as an entry filter | Replay | 2021: +317% became +149% | Failed | Filtered out the best trends |
| Idle cash parked in BTC | Replay | Drawdown up to −54.7% | Failed | More risk, little gain |
| Intrabar real-time exits | Nine-year replay | Final balance −75% vs. bar close | Failed | Noise shakes out winners |
| 1- to 15-minute indicator setups | 17,321 setups that traded, 2024–2026 | 2.7% to 34.2% made money after costs; the 30 that passed the two-period test on 1- and 3-minute bars all did worse than holding the coins | Failed | Costs per trade too large for the moves |
| 4-hour Bollinger breakout 30 / 2.5σ + 2% stop | Replay 2024–2026 and 9 years | +63%/yr in replay after slippage | Paused | Never run live; same backtest caveats |
| 4-hour breakout + trailing exit + BTC filter (method 1) | Walk-forward replay; 2 weeks live paper | Replay +92%/yr; live −3.93%, 5 of 5 lost | Paused | Did not make money live; paused |
"Replay" means a backtest on historical prices with costs. "Paper" means live prices, no real money. Figures in won are paper amounts. Nothing here is investment advice.
The story behind these entries: six weeks of AI trading.