doob

Quantitative strategy research and backtesting CLI. Reads all price data from local ~/market-warehouse/ parquet files — no external API calls for price data.

• Rust 2024 edition (1.85+)
• Populated ~/market-warehouse/ data lake (bronze parquet layer) — see market-data-warehouse

Build and install to ~/.cargo/bin (must be in your $PATH):

cargo build --release
cp target/release/doob ~/.cargo/bin/doob

Verify:

doob list-strategies

After pulling changes or making edits, rebuild and reinstall:

cargo build --release && cp target/release/doob ~/.cargo/bin/doob

# List available strategies and presets
doob list-strategies
doob list-presets
# Run strategies
doob run overnight-drift --no-plots
doob run intraday-drift --ticker SPY
doob run intraday-drift --ticker SPY --short
doob run breadth-washout --universe ndx100 --signal-mode oversold
doob run breadth-ma --universe ndx100 --short-period 50 --threshold 80
doob run breadth-dual-ma --universe ndx100 --short-period 50 --long-period 200 --threshold 20
doob run ndx100-sma-breadth --end-date 2026年03月11日

Build and run the automated candidate-discovery loop from Rust:

# build binaries
cargo build --release
# run with arXiv/Exa web seeding and web-driven net-new candidate exploration
# defaults: --seed-web --candidates 100 --top 10 --verbose --sessions 1008/252
cargo run --release --bin autoresearch_loop -- --seed-web --verbose
# set explicit run envelope and ranking depth
cargo run --release --bin autoresearch_loop -- --seed-web --candidates 100 --top 10 --verbose
# run a larger search (or shorter top list) as needed
cargo run --release --bin autoresearch_loop -- --seed-web --candidates 200 --top 15 --verbose

Optional settings:

• --train-start, --train-end, --test-start, --test-end
• --train-sessions, --test-sessions
• --doob-bin target/release/doob
• --random-seed
• paper-research loop uses paper-research strategy only (no built-in breadth/overnight baseline strategies are queued by default)
• --seed-web uses Exa/arXiv discovery and then generates deterministic mutations around web-proposed ideas

Output artifacts:

• reports/autoresearch-top10-interactive-report.html (interactive top 10 browser report)
• reports/autoresearch-ledger.jsonl (full candidate ledger)
• reports/autoresearch-audits/ (train/test proof JSON for persisted top 10 candidates)
• reports/autoresearch-strategy-registry.json (promoted strategy catalog keyed by stable parameter signature)
• reports/autoresearch-eval-cache.jsonl (cross-run evaluation cache)
• reports/autoresearch-exa-ideas.json (web-seed scrape and normalization)

Output is now table-formatted and includes:

• candidate ID + strategy + strategy category
• focused assets, horizon, source, and rationale
• train/test scores and summary stats
• actual evaluated windows, executed-trade counts, and audit-trail links for persisted winners
• registry promotion history for strong candidates that recur across runs
• machine-copyable command args for the best candidate

To provide the Exa key used for web seeding:

cp .env.example .env
# edit .env: EXA_API_KEY=your_exa_api_key_here

All price data is read from local warehouse parquet files. No Yahoo Finance or external price APIs.

• Price data: ~/market-warehouse/data-lake/bronze/asset_class=equity/symbol=<TICKER>/data.parquet
• Universe membership: presets/<universe>.json (e.g. ndx100.json, sp500.json)
• VIX data: CBOE CSV, cached locally for 24h (only external HTTP call)

A 10-year, 101-ticker backtest with full risk metrics runs in ~0.3 seconds.

All strategies support --output text (default), --output json, and --output md:

doob --output json run overnight-drift --no-vix-filter
doob --output md run breadth-washout --universe ndx100
doob run intraday-drift --ticker SPY --output json

# Unit tests (146 tests)
cargo test
# CLI integration tests (106 tests, requires ~/market-warehouse)
./tests/cli_integration.sh

MIT