Architecture (v2)
How I wired the simplified paper stack. Educational only — not financial advice. Private paths and secrets stay off this page.
0. What I changed
I failed v1 by stacking complexity: regime models, LLM trade proposals, multi-tier risk, options ladders, weekly Sonnet blogs. Cleverness beat compounding.
I start v2 from boring blocks:
- Quality broad-market and dividend ETFs
- Buy weakness (that ticker red today), hold
- Measure everything against my actual paper capital
- Expand only after simple scenarios earn the right
No LLM in the trading loop right now. I want to earn that back later — see Method.
1. Runtime topology
| Piece | Role |
|---|---|
| Main service | python -m trader.main — blocking scheduler |
| Schedule | 10:30 ET and 15:30 ET, Mon–Fri, US RTH only |
| Log viewer | Private FastAPI UI for journal + positions (not this public site) |
Configuration and API keys live in a host-managed environment file — never in git.
| Artifact | Typical relative path |
|---|---|
| SQLite journal | data/logs/trades.db |
| JSONL mirror | data/logs/journal.jsonl |
| Fund universe | data/funds.yaml |
2. Strategy (rule A)
On each scan:
- Load the fixed ETF universe (three sleeves).
- For each ticker, compute its day return (prior close → last).
- Green → skip.
- Red and sleeve underweight vs targets (≈50% / 25% / 25%) → candidate.
- Prefer sleeve primary if red (VOO / SCHD / QQQM); else first red alternate.
- Skip if that sleeve already received a placed buy earlier the same session.
- Market-buy a notional slice of equity (default 5%, cash-capped).
- No auto-sell in software.
Afternoon scan exists so a book that was green at 10:30 can still buy if it turns red later.
schedule (10:30 / 15:30 ET)
→ for each fund
→ day_return(ticker)
→ if green: HOLD
→ if red and sleeve underweight and not bought today: market BUY
→ journal
3. Fund sleeves
Documented universe (primaries first):
| Sleeve | Target | Primary | Alternates |
|---|---|---|---|
| Core | ~50% | VOO | VTI, IVV, SPLG, VT |
| Dividend | ~25% | SCHD | VYM, VIG, DGRO, JEPI |
| Growth | ~25% | QQQM | QQQ, SCHG, VUG, VGT |
Sleeve weight = sum of market value of that sleeve’s tickers ÷ equity.
4. What is intentionally missing
- Regime / Haiku pre-brain
- Claude trader JSON loop
- Options execution and profit ladders
- Soft risk policies and YOLO profiles
- Weekly Sonnet blog generator
- Local LLM / Open WebUI / assistant gateways on the trading host
Those were v1 complexity. Code from that era can live in an archive for archaeology; it is not on the live import path.
5. Journal and operator tools
Every HOLD/BUY writes a journal row (rationale, day return, sleeve, execute status). The private dashboard shows cycles and open positions and links out to the Alpaca paper UI. Public pages here stay educational and redacted.
6. Method going forward
- Keep the live rule set tiny.
- Backtest / paper-validate simple scenarios against the same capital the bot actually uses.
- Add the next knob only when the scoreboard (paper equity + journal) justifies it.
Later (not built): mutate rule variants, score on paper fitness, keep winners. Details on Method.