Auto-Alpha

Autonomous A-share quantitative strategy research powered by Hermes Agent.

Inspired by karpathy/autoresearch: give AI agents a real trading strategy framework and let them experiment autonomously overnight. You wake up to a log of experiments and (hopefully) a better strategy.

How It Works

Nightly cron triggers the orchestrator
  → Researchers modify strategy.py in parallel
    → Evaluator independently backtests
      → Reviewer keeps or discards
        → Writer generates the morning report

Six specialized Hermes profiles collaborate via Kanban. Each has a specific role — no single agent evaluates its own work.

Quick Start

# 1. Install dependencies
uv sync

# 2. Download data (A-share, 5 years)
uv run prepare.py --ticker 000300

# 3. Run baseline evaluation
uv run evaluate.py

# 4. Set up Hermes profiles (see profiles/)
hermes profile create auto-alpha-orch
# ... (repeat for all 6 profiles)

# 5. Start autonomous research
hermes kanban init --board auto-alpha
hermes cron create "every day 23:00" \
  --prompt "Run one auto-alpha research iteration: check program.md for current cycle goals"

Project Structure

auto-alpha/
├── program.md          # Research org charter (human edits this)
├── strategy.py         # Strategy file (agents modify this)
├── evaluate.py         # Evaluation harness (fixed, never modified)
├── prepare.py          # One-time data prep
├── profiles/           # Hermes profile configs (templates, no secrets)
├── skills/             # Agent skills for alpha research workflow
└── research-log/       # Experiment records and reports

Key Design Decisions

1. Fixed evaluation harness — evaluate.py is the single source of truth. Agents cannot modify it, eliminating self-evaluation bias.

2. Composite score — Multi-objective optimization (Sharpe + Calmar + Win Rate) prevents overfitting to a single metric.

3. Train/Val/Test split — 60/20/20 split. Agents only see validation results. Test set is reserved for final evaluation.

4. Independent evaluator profile — A separate agent runs the backtest, so researchers can't inflate their own scores.

5. Kanban audit trail — Every experiment, decision, and report is a durable row in the Kanban board. Full traceability.

License

MIT