DOWTrade

Paper-trading bot for Micro E-mini Dow futures (MYM). A three-LLM analysis pipeline proposes trades; a hard-coded Python safety layer has the final word before any order is placed.

Paper-only. PAPER_ONLY=True and BROKER_ENV="demo" are hard-coded in src/config.py and asserted at startup — the process refuses to run outside a sandbox. No real money, no profitability claims. A learning/research scaffold.

Pipeline (per 15-minute bar)

yfinance ^DJI bar ─▶ feature extraction (ATR-14, SMA-200, Donchian-20, swings)
            │
            ▼
   1. Claude (Anthropic)     — structural read. Sonnet via subscription SDK on
                               crucial bars (position open), else Haiku API
   2. Gemini (API)           — action: open_long | open_short | close | add_pyramid | hold
   3. DeepSeek (HF)          — advisory risk audit (recorded for /disagreements; non-blocking)
            │
            ▼
   4. Python final_check     — HARD override (daily-loss, size, stop, ATR bounds)
   5. Golden/Death cross gate — SMA 20/50 on 15m AND 1hr must agree before entry
            │
            ▼
   sim-fill (paper) ─▶ SQLite ─▶ FastAPI dashboard

The Python layer is the single source of truth for safety. No LLM, env var, or config file can override the rails in src/config.py.

Key features

• Three-LLM ensemble — Claude (structural; Sonnet via the Claude Agent SDK on your Pro plan's included usage for crucial bars, Haiku API otherwise and as fallback), Gemini (execution, via the Gemini API — Google retired the individual-tier gemini-cli on 2026-06-18, so it is off by default), DeepSeek/Qwen (advisory risk audit). Each call is logged. DeepSeek's verdict is recorded for the /disagreements view but no longer gates execution — the deterministic final_check rails are the sole safety authority. With no subscription token everything runs on the metered Haiku API.
• Hard safety layer — final_check enforces max daily loss ($200), fixed risk per trade ($50), max open contracts, mandatory stop-loss, ATR-bounded stops, no averaging down, flat-before-weekend.
• Golden/death cross filter — entries require SMA 20/50 cross agreement on both the 15m and 1hr timeframes (trend-following discipline).
• Sim-fills — SIM_FILLS=1 synthesizes fills locally at bar close and tracks up to 5 concurrent positions, persisted across restarts in sim_state / sim_positions tables (the broker's cert sandbox cannot fill orders).
• Dashboard — FastAPI + Jinja2: equity, day P&L, per-LLM reasoning cards, and a yfinance heartbeat to detect silent stalls.
• Daily journal — APScheduler writes an end-of-day review.

Market data & broker

• Data: yfinance 15m ^DJI (Dow index) bars (polled every 60s), configurable via YF_DATA_SYMBOL / YF_HYDRATE_PERIOD. The index is roll-proof; the previous MYM=F micro-future continuous symbol went dark at the 2026-06-19 quarterly futures roll and stalled the bot, so the structural feed now reads the index (orders are still placed as MYM on the broker). dxLink streaming exists as a fallback but the cert account has no live-data entitlement.
• Broker: Tastytrade certification sandbox. Order submission is stubbed by sim-fills because the sandbox rejects order placement for this account.

Setup

pip install -e ".[dev]"
cp .env.example .env        # fill in API keys + cert credentials

Required in .env: ANTHROPIC_API_KEY, GOOGLE_API_KEY, HUGGINGFACE_API_KEY, and the TASTYTRADE_CERT_* sandbox credentials. See .env.example.

Optional — Claude Sonnet via subscription. To run the structural judge on Claude Sonnet through the Claude Agent SDK (free ~$20/mo subscription credit instead of metered Haiku API tokens), add CLAUDE_CODE_OAUTH_TOKEN (from claude setup-token). Tunables: DOWTRADE_SDK_FOR (positions [default — Sonnet only while holding a live position] | all | none), CLAUDE_SDK_MONTHLY_CAP_USD (default 20), CLAUDE_SDK_MODEL (default sonnet). With no token the bot runs Haiku-only, as before.

Run

python -m src.main          # starts LiveRunner + dashboard on :8000

Tests

python -m pytest                                   # all tests
python -m pytest --cov=src --cov-report=term-missing

Layout

src/
├── config.py            # HARD-CODED safety rails + env Settings
├── main.py              # entrypoint: LiveRunner thread + uvicorn dashboard
├── live/
│   ├── runner.py        # bar loop, LLM pipeline, sim-fills, cross gate
│   └── yfinance_poller.py
├── llm/                 # haiku_structural, gemini_execution, deepseek_risk
│   └── prompts/
├── data/                # bars, features, cross_filter
├── broker/              # tastytrade client, models
├── safety/              # guard layer
├── sizing/              # risk-unit position sizing
├── db/                  # schema.sql + repo (raw sqlite, no ORM)
├── dashboard/           # FastAPI + Jinja2
├── journal/             # daily review (APScheduler)
└── backtest/            # harness: final_check, compute_size, _PositionState
tests/