AI Trading Agent: Macro → Sector → Stock → Technical → Execution

Autonomous regime-aware trading agent with comprehensive multi-layer analysis pipeline.

Macro Analysis (portfolio-level overlay): VIX, yield curve, Fed funds → leverage multiplier (0.5x-1.5x) Sector Analysis (allocation weighting): Sector ETF performance → weights per sector (0.8x-1.2x) Technical Analysis (trade signals): RSI, MACD, support/resistance, Fibonacci, moving averages Strategy Selection (confirmation): Trend, range, breakout, reversal, momentum trading Fundamental Screening (stock quality): P/E, EPS growth, profitability, debt levels Regime Detection (HMM): Market state (crash/bear/neutral/bull/euphoria) → base allocation Execution (orders): Position sizing × macro × sector × technical → limit order placement Safety Limits (circuit breakers): 2% daily loss, 5% max drawdown → throttling/halt

Setup

1. Create .env file from template:

   cp .env.example .env

   Add your Alpaca paper trading API credentials.

2. Install dependencies:

   source venv/bin/activate
   pip install -r requirements.txt

3. Initialize system:

   python scripts/orchestrate.py init

   Verifies API connection, market status, configurations.

4. Configure Claude Code routines:

   • The .claude/routines.json file defines three scheduled tasks
   • 9:45 AM: Morning research (regime detection + market data)
   • 10:00 AM: Trading session (allocation + execution)
   • 4:15 PM: End-of-day journal (finalize logs)

Files

Multi-Layer Analysis:

• scripts/macro.py — VIX, yield curve, Fed funds → leverage multiplier (0.5x-1.5x)
• scripts/sector.py (in fundamentals.py) — Sector ETF returns → allocation weights
• scripts/technical.py — RSI, MACD, support/resistance, Fibonacci, moving averages
• scripts/strategies.py — Strategy selection (trend, range, breakout, reversal, momentum)
• scripts/fundamentals.py — Stock screening (P/E, EPS, profitability, debt, ROE)

Core Engine:

• scripts/regime.py — HMM regime detection (crash/bear/neutral/bull/euphoria)
• scripts/allocation.py — Position sizing (regime × volatility × trend × macro × sector × technical)
• scripts/safety.py — Circuit breakers (2% daily loss, 5% drawdown)
• scripts/backtest.py — Walk-forward backtesting
• scripts/orchestrate.py — Main orchestration (research → trading → EOD)

Data & Trading:

• scripts/research.py — Market data fetching (bars, news, account)
• scripts/trade.py — Order placement + market status checks

Configuration:

• CLAUDE.md — Agent instructions, regime rules, safety limits
• watchlist.json — Symbols, allocation limits per position
• .claude/routines.json — Scheduled task config
• .claude/settings.json — Claude Code project settings

Monitoring:

• dashboard.py — Streamlit UI (positions, regime, safety, journal)
• journal/ — Trade logs, regime history, safety state

Running

System initialization:

python scripts/orchestrate.py init

Individual routines (manual test):

python scripts/orchestrate.py research
python scripts/orchestrate.py trading
python scripts/orchestrate.py eod

Main execution loop (5-minute bars):

python scripts/orchestrate.py loop 300

Runs continuously, checks regime/safety every 5 min.

Data tools (debugging):

python scripts/research.py account
python scripts/research.py bars AAPL
python scripts/trade.py status

Dashboard:

streamlit run dashboard.py

Opens UI at localhost:8501 showing positions, regime, safety status, trade journal.

Market Regimes (HMM-Detected)

Regime	Signal	Base Leverage	Cash Reserve
Crash	↓ High vol, downtrend	0.0x	95%
Bear	↓ Medium vol, downtrend	0.5x	70%
Neutral	↔ Low vol, sideways	1.0x	20%
Bull	↑ Medium vol, uptrend	1.5x	15%
Euphoria	↑ High vol, unsustainable	1.0x	30%

Detected via HMM on 20-day returns + intraday volatility (3-7 auto-selected states).

Trading Strategies (Adaptive)

Agent selects strategy based on market conditions:

Strategy	Trigger	Entry	Exit	Best For
Trend Trading	Bull + strong uptrend	Dips to MA support	Trend break	20%+ sustained moves
Breakout Trading	Consolidation + low vol	Close above resistance	Below breakout	Volatility expansion
Range Trading	Sideways (neutral)	At support	At resistance	Choppy, low-vol markets
Reversal Trading	Bottom/top reversal detected	Confirm reversal pattern	Opposite reversal	Mean-reverting bounces
Momentum Trading	ROC >1% despite pullback	Dips in strong trend	Overbought (RSI >70)	Add to winning positions

Technical confirmation (RSI <70 for buy, >30 for sell) + macro overlay (reduce size in bearish macro).

Macro Overlay (Portfolio-Level)

Adjusts leverage based on macro conditions:

Condition	Leverage Multiplier	Rationale
VIX >30, inverted curve, high rates	0.5x	Risk-off environment
VIX 15-20, normal curve, moderate rates	1.0x	Neutral macro
VIX <15, normal curve, stable rates	1.5x	Risk-on environment

Applied to all positions equally. Reduces risk in bearish macro, increases in bullish.

Safety Rules (Hardcoded)

• Max daily loss: 2% of opening value → position throttle activates
• Max drawdown: 5% from peak → new trades limited
• Stop loss: 8% below entry → force close
• Circuit breaker: Both limits breach → halt trading
• Error resilience: >5 consecutive errors → shutdown (prevents cascading failures)

Important

• Paper trading only — Start with APCA_BASE_URL=https://paper-api.alpaca.markets
• Live trading prep — Validate 2+ weeks of paper trading before switching to live keys
• Regime-based sizing — Position limits adjust dynamically, not static 5%
• All decisions logged — journal/YYYY-MM-DD.md contains full reasoning chain
• State persistence — journal/orchestrator_state.json tracks errors, enables recovery

Monitoring

• Daily logs: journal/YYYY-MM-DD.md — research, trades, EOD summary
• Regime history: journal/regime_history.json — regime + confidence over time
• Safety state: journal/safety_state.json — peak capital, daily P&L, throttle
• Error tracking: journal/orchestrator_state.json — error count, last successful run
• Dashboard: Streamlit UI updates every 60 seconds

Backtesting

Walk-forward test (avoids look-ahead bias):

python scripts/backtest.py

Trains on 60 days, tests on 20 days, rolling window. Returns Sharpe ratio, total return, drawdown.

Troubleshooting

API connection fails:

python scripts/research.py account

If this fails, check .env credentials and internet connection.

Market closed error: Market-dependent routines skip automatically. Check next open time in status output.

Circuit breaker active: Check journal/safety_state.json for daily loss % and drawdown %. Wait until reset or reduce portfolio exposure.

Component sync fails: Run python scripts/orchestrate.py init to reinitialize all systems.

Costs

• Claude inference: ~$5-10/month for 3 daily sessions
• Alpaca: Free for stocks, some crypto fees
• Hosting (if remote): ~$5-10/month for light VPS

Streamlit dashboard runs locally or free at streamlit.io.