🤖 Multi-Agent Quantitative Trading System

A macro buy-and-hold engine that orchestrates a team of specialized analysis agents — technical, trend, macro, sentiment, and quantitative alpha — into a single 0–100 conviction score, then sizes and executes a live portfolio through Alpaca.

📈 Live Paper-Trading Performance

Real money-weighted Alpaca paper account, auto-generated from the broker API.

Account opened 2026-03-13. Snapshot 2026-06-04. Past performance ≠ future results.

| | Return (since inception) | |---|---:| | 🤖 This strategy (paper) | +53.6% | | S&P 500 (SPY) | +13.5% | | Nasdaq-100 (QQQ) | +24.8% |

Starting capital $100,000 → peak equity $159,560, max drawdown −8.4%, 32 open positions concentrated in semis/AI (MRVL +33%, ARM +18%, SNDK +12%).

📊 Full breakdown + equity curve → docs/PAPERTRADING.md (refresh anytime with python scripts/fetchpaper_performance.py)

🧪 Backtest — Strategy vs S&P 500 & Nasdaq-100

Per-calendar-year, point-in-time backtest with T+1 execution and 5 bps/side transaction costs (the same weighting code the live bot trades):

| Year | 🤖 Strategy | S&P 500 | Nasdaq-100 | |------|------------:|--------:|-----------:| | 2022 | −45.4% | −18.6% | −33.2% | | 2023 | +92.4% | +26.7% | +55.9% | | 2024 | +60.1% | +25.6% | +27.7% | | 2025 | +24.2% | +18.0% | +21.0% | | 2026 (YTD) | +26.6% | +10.7% | +21.5% |

⚠️ Honesty note: the universe is today's survivors held back through time (survivorship bias), so treat the raw alpha as an upper bound. The strategy is a high-beta amplifier — it shines in up years and overshoots drawdowns in down years (see 2022). The full, self-critical analysis is in docs/BACKTEST_RESULTS.md — including the strategy's edge over its own basket, which is small and noisy. We publish the warts on purpose.

🧠 Multi-Agent Architecture

The score for every stock is produced by a panel of independent analysis agents, each looking at the market through a different lens, then fused by a weighted scorer:

┌─────────────────────────────┐
   Market data ─────────▶│        DATA LAYER           │  yfinance: 2y daily + 1y hourly
   (prices, news, SPY)   │  data_fetcher / discovery   │
                         └──────────────┬──────────────┘
                                        ▼
        ┌───────────────────────────────────────────────────────────────┐
        │                     ANALYSIS AGENTS                            │
        ├───────────────┬───────────────┬──────────────┬────────────────┤
        │ 📐 Technical  │ 📊 Trend      │ 🌍 Macro     │ 📰 Sentiment   │
        │ SMA/RSI/MACD  │ rel-strength  │ regime &     │ news headline  │
        │ ADX/Boll/OBV  │ vs SPY        │ risk on/off  │ lexicon+decay  │
        │   (35%)       │   (30%)       │              │   (15%)        │
        ├───────────────┴───────────────┴──────────────┴────────────────┤
        │           🔒 Quantitative Alpha — 30 factors (20%)             │  ← core/ (private)
        └───────────────────────────────┬───────────────────────────────┘
                                        ▼
                         ┌─────────────────────────────┐
                         │   🔒 SCORER  →  0–100 score │  confidence- & regime-adjusted
                         │   🔒 STRATEGY → target wts  │  top-10 momentum, max 25%/name
                         └──────────────┬──────────────┘
                                        ▼
                         ┌─────────────────────────────┐
                         │  Alpaca executor + report   │  T+1 rebalance, email digest
                         └─────────────────────────────┘

Plus a layer of Claude research agents (in .claude/skills/) that enrich the macro view — market-news-analyst, scenario-analyzer, market-environment-analysis, earnings-calendar, and economic-calendar-fetcher.

🔒 The proprietary signal engine — the 30 alpha factors, the scorer, and the position-sizing strategy — lives in a private core/ package that is not included in this repository. Everything else (the harness, analyzers, backtester, executor) is open.

🎯 Scoring Model

| Agent | Weight | Signals | |-------|:------:|---------| | 📐 Technical | 35% | SMA 50/200, RSI, MACD, ADX, Bollinger, OBV | | 📊 Trend | 30% | Relative strength vs SPY, market regime | | 🔒 Alpha | 20% | 30 quantitative factors (WorldQuant-style) | | 📰 Sentiment | 15% | News headline analysis with recency decay |

Scores are confidence-adjusted and regime-aware (conservative in bear markets, a slight boost in bull markets).

| Score | Recommendation | Grade | |:-----:|----------------|:-----:| | 75+ | Strong Buy | A | | 60–74 | Buy | B / B+ | | 45–59 | Hold | C / C+ | | 30–44 | Reduce | D | | <30 | Avoid | F |

🚀 Quick Start

The private core/ package is required to run end-to-end. Without it the harness imports the scorer/strategy/alpha modules from core/ and will fail — by design, the alpha is not published.

python -m venv .venvtrading && source .venvtrading/bin/activate
pip install -r requirements.txt
cp .env.example .env            # then fill in your Alpaca keys
Analyze the full universe (157 US equities)
python main.py
Specific tickers, top 10, skip news for speed
python main.py --tickers AAPL,NVDA,MSFT --top 10 --quick
Daily report + paper trade
python daily_report.py --trade
Backtest vs SPY (a single recent year, verbose)
python backtest_strategy.py --tickers AAPL,MSFT,NVDA,GOOGL,AMZN,META,AVGO,TSLA,AMD,CRM,ORCL,ADBE
Multi-year sweep vs SPY & Nasdaq
python backtest_years.py

CLI options (`main.py`)

--tickers AAPL,MSFT    Analyze specific stocks (default: full universe)
--top N                Show top N results only
--quick                Skip news sentiment (faster)
--no-alpha             Skip alpha factor computation (faster)
--output-dir DIR       Report output directory (default: reports/)
--no-report            Console output only, skip report file

📁 Repository Layout

main.py  daily_report.py        Entry points (CLI + daily runner)
data_fetcher.py                 Prices, news, benchmarks (yfinance)
technical_analyzer.py           Trend / momentum / volume / volatility
trend_analyzer.py               Relative strength vs SPY, market regime
macro_analyzer.py               Market regime & risk-on/off
sentiment_analyzer.py           News headline sentiment
stock_discovery.py              Weekly universe expansion
report_generator.py             Markdown report output
email_sender.py                 Email digest
alpaca_trader.py                Paper/live execution via Alpaca
backtest_strategy.py            Strategy vs buy-and-hold (single window)
backtest_years.py               Per-year sweep vs SPY & Nasdaq
core/             🔒 PRIVATE    custom_alphas · scorer · strategy · configs
docs/                           STRATEGY · BACKTESTRESULTS · PAPERTRADING
scripts/                        fetchpaperperformance.py
.claude/skills/                 Claude research agents

⚙️ Deployment

Local daily agent (macOS): ./setupdaily.sh install renders a git-ignored launchd plist from *.plist.template and schedules dailyreport.py for 8 AM on trading days.
Container: docker build -t trading-bot . — secrets are injected at runtime (--env), never baked into the image.

🔐 Security & Privacy

Secrets live only in .env (git-ignored). No API keys, passwords, or personal emails are tracked — see .env.example for the required variables.
The launchd plist is rendered locally and git-ignored; only the placeholder *.plist.template is tracked.
The proprietary core/ algorithm is git-ignored and absent from this public repo.

⚠️ Disclaimer

This project is for research and educational purposes only. It is not investment advice. Markets carry risk; past and backtested performance does not guarantee future results. Trade live money at your own risk.