An autonomous derivatives-strategy research agent — think of it as an ML intern that researches strategies for you. Given a natural-language task it runs the full research loop itself (fetch data → analyze → backtest → read metrics → iterate → honest verdict) and proves every claim with rigorous out-of-sample backtesting instead of asking you to trust a black box.
It has two isolated research desks sharing one agent core (separate prompt, playbook, tools, and conversation context — they never mix):
| Desk | --profile |
Universe | Data (public/free first) | What it researches |
|---|---|---|---|---|
| 股票期權 | options |
US equity options | ThetaData (free) / Databento (paid) + yfinance | VRP screening, cash-secured-put / covered-call / wheel backtests, walk-forward, unusual activity, charts, support/resistance |
| 幣安永續 | crypto |
Binance USDⓈ-M perps (incl. US-stock perps like NOKUSDT/AAPLUSDT) | fapi.binance.com public REST (no key) |
funding-rate carry (the perp analog of VRP), grid-bot backtest |
The perp desk uses public market data only — no API key, no order placement: it analyzes and backtests, it does not trade.
📐 Full architecture (layered diagram, agent-loop sequence, per-desk data flow, module reference): docs/architecture.md.
OptionPilot stands on two pillars:
-
An ml-intern-style autonomous research loop. Give it a task — "research cash-secured puts on ZETA" or "is a grid bot worth running on NOKUSDT?" — and it runs the experiment loop itself: propose a hypothesis → fetch data → analyze → backtest → read metrics → propose the next improvement → iterate → deliver the best strategy with a full, honest experiment log. It proposes, iterates, and records on its own (
ExperimentLoop+Planner+ExperimentTracker). -
A rigorous, verifiable backtesting system. Every strategy must pass the same honest bar: out-of-sample / walk-forward evaluation, realistic costs (spread, slippage, fees, assignment, funding), full metrics (Sharpe, max drawdown, win rate, turnover) plus worst-case stress tests, reproducible head-to-head run comparison, and honest reporting of negative results.
The same methodology runs on both desks. The options desk screens variance risk premium and backtests put/call selling; the perp desk treats the funding rate as the structural analog of VRP and stress-tests grid bots — separating booked grid profit from stuck-inventory loss, and showing how a grid bleeds in a trend. Both always benchmark against simply buying & holding.
Success is not "find a magic money signal" — it's being able to answer, for any candidate strategy, how much edge actually survives costs out-of-sample, and what the worst case is, so capital only goes to strategies that pass the bar.
Architecture borrows patterns from huggingface/ml-intern (agentic loop, ContextManager, ToolRouter, approval gating, doom-loop detection) but is written fresh with no Hugging Face coupling, specialized for options & perpetual-futures research.
Early scaffold. See plan.md for the full design and roadmap.
| Area | Choice |
|---|---|
| Options data | ThetaData free tier (local terminal, recent ~2yr, real bid/ask + volume) as default; Databento OPRA (deep history, pay-as-you-go) for older tickers; greeks computed locally |
| Perp data | Binance USDⓈ-M public REST (fapi.binance.com) — klines, funding rate, open interest, long/short ratio; no API key |
| Desk isolation | two profiles (options / crypto) with separate prompt, playbook, tools, and context — agent/profiles.py |
| Agent harness | written fresh, borrowing ml-intern patterns; multi-model via LiteLLM |
| Validation | out-of-sample / walk-forward + realistic costs (spread, slippage, fees, funding) before any capital is risked |
| Cost control | fetch-size/cost guard + approval gating on expensive (Databento) pulls |
Full setup (package + LLM + data sources) is in docs/INSTALL.md.
uv sync --extra dev --extra data # install
cp .env.example .env # set OPTIONPILOT_DATA_SOURCE + keys (see INSTALL.md)
# options desk (default)
uv run optionpilot "回測 ZETA 近一年的賣 put 跟持股比"
# perp desk — public Binance data, no key needed
uv run optionpilot --profile crypto "分析 NOKUSDT 的資金費率 carry，並回測網格"
uv run optionpilot # interactive (add --profile crypto for the perp desk)A Gradio app with two tabs (股票期權 / 幣安永續), each its own isolated session, streaming the agent's tool calls + verdict live and showing charts inline:
uv sync --extra ui
bash scripts/serve_local.sh # local model must be running (separate terminal)
uv run optionpilot-ui # opens http://localhost:7860Run a local OpenAI-compatible server and point OptionPilot at it — no API key, no rate limits.
bash scripts/serve_local.sh # vLLM serving Qwen3-Coder-30B-A3B (FP8) on one GPUThen in .env:
OPTIONPILOT_MODEL=hosted_vllm/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8
OPTIONPILOT_API_BASE=http://localhost:8000/v1
OPTIONPILOT_API_KEY=local
Caches are kept on /media/user/data2 (the root disk is small); see the script header.
See docs/architecture.md for the full picture; in brief:
optionpilot/
cli.py Typer entry (interactive / headless, --profile options|crypto)
config.py config hierarchy (env > defaults)
ui/app.py Gradio GUI — two isolated desks as tabs
llm/ LiteLLM client wrapper (local vLLM / cloud)
agent/ loop, context, router, doom-loop, approval, planner,
playbook (options + crypto), profiles (desk isolation), lang (繁體)
tools/ options: measure_vrp, run_backtest, optimize_strategy,
detect_unusual_activity, make_charts, support_resistance, fetch_options_data
crypto: funding_analysis, grid_backtest
shared: ask_user, list_experiments
analysis.py VRP (implied vs realized, up/down split) + support/resistance
crypto.py perp funding summary (carry) + realized vol from klines
data/ sources (ThetaData/Databento ABC), market loaders, databento fetcher
(+ cost guard), osi parser, binance public client, Black-Scholes greeks
backtest/ strategies (CSP/CC/wheel), grid (perp grid bot), walkforward, metrics
signals/ unusual options activity + put/call flow
plots.py matplotlib charts (CJK 繁體)
tracking/ experiment tracker (DuckDB) + Markdown reports