Robin

Robin is a session-native, host-orchestrated agentic quant research platform. It generates factor hypotheses, critiques them, implements math or deep-learning factor panels, validates those panels, converts the best panels into executable portfolio strategy candidates, and promotes only strategies that pass out-of-sample gates.

This repository is a cleaned public extraction and refactor of the multi-agent quant work originally prototyped inside the autotrade project.

Research only. This project is not financial advice and does not place trades.

What It Does

• Research Agent proposes traditional formula factors and optional PyTorch sequence factors.
• Debate Agent writes bullish, bearish, and conservative validation cases before implementation.
• Implementation Agent compiles safe JSON-style feature expressions or trains compact sequence models.
• Validation Agent evaluates IC, rank IC, OOS rank IC, decile spread, Sharpe, drawdown, coverage, and stability.
• Fusion step combines factor panels with equal-weight, IC-weighted, or stacking logic.
• Strategy Agent builds top-k rotation strategies from accepted/watch factors.
• Portfolio Backtest Agent evaluates OOS return, OOS Sharpe, drawdown, turnover, exposure, and benchmark-relative excess.
• Memory layer writes JSONL experiments, checkpoints, and Markdown knowledge updates.
• Session Host wraps each run in an auditable research session with isolated events, transcripts, checkpoints, and artifacts.

Install

git clone <your-fork-url> robin
cd robin
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

For deep-learning factors:

pip install -e ".[dev,dl]"

Without PyTorch, DL factor attempts fail gracefully and the loop still evaluates traditional factors.

Quick Start

Run a deterministic offline cycle:

robin --offline-synthetic --max-iterations 1 --universe-size 24

Or without installing the console script:

PYTHONPATH=src python3 -m open_quant_agent.cli --offline-synthetic --max-iterations 1

Run with yfinance data:

robin --max-iterations 1 --period 3y --interval 1d

The legacy command name remains available:

open-quant-agent --offline-synthetic --max-iterations 1

Sessions

Create, run, and inspect an isolated research session:

robin session create --goal "Find robust semiconductor supply-chain alpha" --offline-synthetic --no-network
robin session run qrs_xxx --max-iterations 1 --offline-synthetic
robin session inspect qrs_xxx

Outputs

The legacy loop writes local research artifacts:

• .agent_state/multi_agent_quant_state.json - supervisor arm state plus factor and strategy history.
• .agent_state/checkpoints/cycle_XXXX.json - full cycle checkpoints.
• multi_agent_experiments.jsonl - machine-readable experiment memory.
• EXPERIMENTS_LOG.md - cycle summaries.
• RESULTS.md - factor and strategy result summaries.
• FACTOR_KNOWLEDGE_BASE.md - factor memory.
• STRATEGY_KNOWLEDGE_BASE.md - strategy memory.

These files are ignored by git by default.

Session runs write isolated artifacts under:

• runs/sessions/{session_id}/session.json
• runs/sessions/{session_id}/status.json
• runs/sessions/{session_id}/events.jsonl
• runs/sessions/{session_id}/transcript.jsonl
• runs/sessions/{session_id}/.agent_state/checkpoints/
• runs/sessions/{session_id}/multi_agent_experiments.jsonl

Strategy Promotion

The strategy layer intentionally separates "good absolute backtest" from "production-worthy candidate." A strategy is accepted only when it has positive OOS return, positive OOS excess versus the equal-weight benchmark, controlled drawdown, enough rebalances, and enough exposure.

That means a strategy can show strong OOS Sharpe but remain on watch if it still loses to the benchmark.

Architecture

See:

• docs/architecture.md
• docs/agent_roles.md
• docs/workflow.md

Daemon Mode

For a simple local research daemon:

scripts/run_daemon.sh

Set cadence with:

ROBIN_SLEEP_SECONDS=900 scripts/run_daemon.sh

Tests

pytest

License

MIT. See LICENSE.

Attribution

Robin is derived from and substantially refactored out of the local autotrade research project. See NOTICE.