This agent monitors four independent data sources for trading signals. When 2+ signals from different sources converge on the same ticker within 7 days, an LLM synthesizes them into a holistic trading decision.
| Source | Data | Edge | Frequency |
|---|---|---|---|
| 8-K | SEC EDGAR material event filings | LLM reads filings faster than human analysts | Every 30 min |
| EARNINGS | Yahoo Finance earnings estimates vs actuals | PEAD: stocks drift 30-60 days after surprise | Every 30 min |
| INSIDER | SEC EDGAR Form 4 insider purchases | 2+ officers buying = strong conviction | Every 30 min |
| TECH | Price, volume, RSI (Yahoo Finance) | Confirms momentum — not predictive alone | Every 30 min |
Each signal fires independently into a rolling 7-day ledger. When 2+ signals from different sources appear on the same ticker, the LLM brain (Claude Sonnet) analyzes them holistically and assigns a conviction score (1-10). Only conviction ≥ 7 triggers a trade.
| Conviction | % of Portfolio | Frequency |
|---|---|---|
| 7 | 3% | Regular (good signals) |
| 8 | 5% | Strong (clear catalyst + confirmation) |
| 9 | 7% | Very strong (rare, exceptional) |
| 10 | 10% | Maximum (once per quarter caliber) |
| Rule | Trigger | Action |
|---|---|---|
| Hard Stop | Price drops 15% below entry | GTC stop order on Alpaca |
| Trailing Stop | Price rises 12%+ above entry | Cancel hard stop, trail 10% below peak |
| Time Stop | 30 trading days (~42 calendar) | Market sell at close |
Max 12 concurrent positions. Max 25% in any one sector. If portfolio drops 5% in a single day, all new entries paused for 24 hours. No averaging down — ever.
Uses Anthropic Claude API (undefined) with temperature 0.3. Max 100 calls/day, 24-hour result cache. Estimated cost: ~$0.50-2.00/day.