BTC Day Trading

Fully automated AI trading system. No human intervention — AI decides entries, exits, and position management.

Exchange: Hyperliquid DEX (testnet), BTC-USDT perp, 40x isolated leverage
AI Agent: Kiro CLI (Claude) — reads market data, builds theories, executes trades via Python SDK
Cycle: Full analysis → execution → 6 check-ins every 2h (~12h per cycle, 2 cycles/day)
Position size: ~$4,000 notional (~$100 margin per trade)

Data fed to AI each cycle:

6 OKX chart screenshots (1m, 5m, 15m, 1H, 4H, 1D) — candles, EMAs, BB, Ichimoku, OI, CVD, RSI, MACD, ADX, MFI, Stoch, ATR — view
6 per-timeframe JSON reports — 20 candles with full indicator history, CVD buy/sell VWAP, candlestick patterns — view
Analysis JSON — S/R levels, funding rate history, OI, 24h stats — view
Hyperliquid positioning — cohort breakdown (whales to shrimp), bias trends, liquidation heatmap — view

Apr 28 – May 4: AI-only strategy (theory-first analysis → execution → check-ins). Result: $1,000 → $640 (−36%). AI traded like the crowd — tight SLs got hunted, panic-closed positions, even disabled its own cron once.
May 5 – May 9: Added "Inverse Bot" — second wallet doing exact opposite of main AI. Theory: if AI = crowd, inverting = trading with MMs. Result: both lost. Inverse $999 → $837 (−16%). Problem: inverse closed when main closed (killing its own profits), bad R:R from mirroring tight SLs as TPs, slippage on delayed entry.
May 9: Inverse disabled. Key lessons: mechanical inversion doesn't work because (1) AI doesn't purely predict crowd — it's sometimes right, (2) timing lag kills edge, (3) "close when main closes" rule destroyed the one advantage inverse had.
Current: Contrarian approach — AI analyzes market, then weighs that its "obvious" reads historically matched retail thinking. No stop losses, liquidation as stop. Let's see.