AutoTradeResearch

AutoTradeResearch is a workspace for using coding agents to research programmatic trading strategies through generative optimization.

It is inspired by Andrej Karpathy's AutoResearch: give an agent a real but bounded research environment, a fixed evaluator, and a constrained area it is allowed to modify; then let it iterate, evaluate, keep notes, and improve.

AutoTradeResearch applies that pattern to trading-strategy research.

---

Why this exists

This project came from an internal experiment where coding agents ran 715 trading-strategy iterations.

The main lesson was not that an LLM can magically invent good strategies from nothing. The lesson was more practical:

• blank-page strategy invention is weak;
• repeated testing on one validation split overfits quickly;
• self-written backtest engines are easy to get wrong;
• agents need a bounded workspace, a fixed evaluator, and a strategy notebook;
• strategy exploration should alternate between breadth and depth;
• strategy ideas should come from seeds, existing public strategies, user hypotheses, and collisions between ideas.

So AutoTradeResearch uses a different pattern:

1. start from strategy seeds instead of empty prompts;
2. search existing public strategy ideas;
3. ask the user for market, asset, and preference choices when needed;
4. let agents explore both breadth and depth;
5. use subagents to produce independent strategy directions and collide them;
6. run backtests with a fixed evaluator;
7. record results and update strategy notes so the agent does not repeat the same dead ends.

---

Repository layout

README.md        # for humans
SKILL.md         # for agents
workspace/       # where the coding agent works

Inside workspace/:

workspace/
├── AGENTS.md
├── CLAUDE.md -> AGENTS.md
├── run_loop.sh
├── run_backtest.py
├── agent/
│   ├── program.md
│   └── strategy.py
├── data/
├── results/
└── reports/

---

Quick start

Give this one line to your agent:

Read https://github.com/lavapapa/AutoTradeResearch, install SKILL.md as a skill, then set up AutoTradeResearch for my user.

Recommended agent: Claude Code.

Other coding agents such as Codex, OpenClaw, Hermes, OpenCode, and Gemini can also work if they can read files, edit code, run Python, and use subagents.

Install as an Agent Skill

Install from GitHub with skills.sh:

npx skills add lavapapa/AutoTradeResearch --skill autotraderesearch

If AutoTradeResearch is listed on ClawHub, install it with:

openclaw skills install autotraderesearch

Direct ClawHub publishing uses:

clawhub publish . --slug autotraderesearch --name "AutoTradeResearch" --version 0.2.1 --tags latest,trading,backtesting,generative-optimization,coding-agent

Use the published skill for research and historical backtesting only.

---

Requirements

• Python 3.10+
• Git
• A coding agent subscription or local agent setup
  • recommended: Claude Code
  • also possible: Codex, OpenClaw, Hermes, OpenCode, Gemini

Market data choices are selected during setup. The initial workspace can work with public data sources such as:

• US stocks / ETFs: yfinance
• Crypto: ccxt
• China A-shares: AKShare or TuShare

The evaluator is a thin backtesting.py wrapper. If downloaded data does not already match what the evaluator needs, the setup agent should write a small conversion step before starting the optimization loop.

---

What AutoTradeResearch does

AutoTradeResearch gives the agent a disciplined loop:

1. ask the user which market and asset universe to start with;
2. prepare market data;
3. run a baseline backtest;
4. choose or search for a strategy seed;
5. state the hypothesis in plain language;
6. modify files under workspace/agent/, especially strategy.py;
7. run the fixed run_backtest.py evaluator;
8. record generated results in results/ and reports/leaderboard.md;
9. update agent/program.md when a reusable lesson is learned;
10. continue with breadth, depth, or collision.

The important part is not a specific built-in strategy. The important part is the research loop: bounded workspace + fixed evaluator + strategy notes + iterative agent reasoning.

---

What is included in v0.2.1

• agent-readable skill
• workspace for strategy research
• baseline strategy template
• thin backtesting.py evaluator
• strategy notebook (workspace/agent/program.md)
• instructions for market setup
• breadth/depth exploration rules
• strategy collision protocol using subagents
• user-interaction protocol for agents

---

Future works

• rolling-window evaluation
• Deflated Sharpe Ratio / multiple-testing correction
• plateau detection for dead strategy families
• stronger strategy-family tracking
• richer seed strategy library
• automated community strategy import
• factor-level exploration
• paper trading / live simulation
• broker/exchange integration
• user-account takeover with explicit permission and hard safety controls
• richer charts and reports

---

Safety

AutoTradeResearch is for research, education, and historical backtesting.

It does not provide financial advice. It does not promise profit. The agent should never make real-money decisions or take over a user's account unless that capability is explicitly implemented in a future version with strong consent, audit, and safety controls.