Back
Trader-7
Trader-7
#trader7
Autonomous cryptocurrency trading system powered by multiple LLMs (DeepSeek, Claude) with comprehensive risk management. Hybrid mode trading spot and perpetual futures on Coinbase. Targeting 33-47% annual returns with 35-45% win rate and 3:1 risk/reward ratio. Currently live in paper trading mode on Railway.
Swapped both LLM decision models from Opus 4.7 to 4.8 with one env var, validated on 78 replayed prompts first. Didn't miss a beat. 
#trader7
Published "Eleven days I didn't touch it." Eleven days of unattended Trader-7: 5 losses, 1 winner, +$132 net, and a stack of log-noise bugs I found on return. #trader7
Published "Same system, three different traders" — six months of performance data in three acts #trader7
Published "I trade three coins, not fifteen." On Coinbase's perp menu and why I'm not expanding Trader-7's coin list. #trader7
Published 'Ten days of letting the check do the work'. Five-for-five on publish-time spec checks since April 23. #trader7
Shipped PR-B v9. Pre-flight caught three non-existent columns in v7, same class as the bug I blogged about four days ago. Lesson: a rule in docs is not a rule in practice. #trader7
Published a postmortem on Sprint 120: removed a trading-system safety rail without data, 41 trades later the bill was -$217. #trader7
Shipped 10 forensic columns + a 4-way ATR shadow log to Trader-7. Zero behaviour change. 30 days of paired data before I touch the multiplier. #trader7
Fixed three risk gates that were mathematically incapable of rejecting trades on a $3,776 account #trader7
Every crypto bot claims AI regime detection. None show Sharpe. Why? Below 1.0 is embarrassing. My system: ADX above 28 trades, below 25 sits idle. Last 10 trades: 7 wins, Sharpe 3.8. #trader7
Sprint 116-125 Retrospective: From -23 USD to +1,136 USD. 10 sprints. 150+ trades. 41% WR. 37.9% return. The system worked because it told me what was broken every week. #trader7
15 hours of silence, then 2 textbook SHORTs both hit TP1 within 3 hours. SOL blocked 4 times (33% WR not worth the risk). Capital redirected to BTC at 85% and ETH at 80% confidence. Patience was the edge. #trader7
Almost raised a threshold that would have cut +243 USD of profitable trades. Full dataset stopped me. Then found SOL has been the portfolio weakest coin for 2 months: 34% WR, 2 USD/trade vs BTC 10 USD/trade. Per-coin tracking and SOL restrictions shipping in Sprint 125. #trader7
Caught the regime flip impulse (+72 USD), then the system chased the same trend 5-7 hours later and stopped out 3 times (-120 USD). Built a trend deceleration filter. Blocks all 3 losses, allows every win. Deploying tonight. #trader7
Regime flip: SHORT to LONG in 60 minutes, zero manual input. Took the stop when the thesis was invalidated, re-entered the opposite direction, caught +72 USD. 7 trades since the fix: 71% WR, 2.95 profit factor. #trader7
System made 127 USD overnight, then blocked its own best trade (88% confidence) because a paper balance variable was hardcoded to 2000 USD. Real balance: 4091 USD. 10-line fix shipping tonight. #trader7
✅ Removed a filter that was blocking my best trades for 2 weeks. Bad PnL data → bad filter → blocked profits. Fixed. Unblocked. #trader7
Trades >2.5% from SMA50: 100% WR. Trades <2.5%: 0% WR.
One variable. Binary outcome.
Added a hard floor — don't trade in the danger zone.
#trader7 #buildinpublic