Inspired by DeepMind’s AlphaEvolve coding agent—this project applies the same evolutionary-LLM principles to financial markets.
Autonomously discovers and back‑tests high‑performing algorithmic‑trading strategies using evolutionary LLM prompts, Backtrader, and the Papers‑With‑Backtest data ecosystem.
# Clone and install in editable mode
$ git clone https://github.com/paperswithbacktest/pwb-alphaevolve.git
$ cd pwb-alphaevolve
$ pip install -e .
# Set your OpenAI key (model "o3" required)
$ export OPENAI_API_KEY=sk-...
# Set your Papers‑With‑Backtest dataset (e.g. "paperswithbacktest/Stocks-Daily-Price")
$ export HF_ACCESS_TOKEN=hf_Seed strategy code can be found on paperswithbacktest.com and the required data for the backtest can be obtained from paperswithbacktest.com/datasets.
Launch and monitor an experiment entirely from the GUI:
$ streamlit run scripts/gui.pyThe sidebar lets you paste a seed strategy, tune the options from
examples/config.py and pick the number of iterations. Click Run evolution to start
the search and watch the hall‑of‑fame table update live.
Alternatively you can launch the evolution controller (create a fresh experiment)
python scripts/run_example.py --experiment my_expUse the --experiment option to keep runs separate:
python scripts/run_example.py --experiment my_expThis creates a new SQLite file my_exp.db under ~/.alphaevolve/. The GUI lists all experiments, allowing you to switch between them or delete one via the Delete experiment sidebar button.
Python ≥ 3.10 required.
pip install pwb-alphaevolveOr install the bleeding‑edge version:
pip install git+https://github.com/paperswithbacktest/pwb-alphaevolve.git- pwb-toolbox
- pwb-backtrader
- openai ≥ 1.0 (structured output)
- tqdm, pandas, numpy, pydantic
(See pyproject.toml for the full list.)
AlphaEvolve can operate without OpenAI by loading a HuggingFace model or by forwarding requests to an OpenAI-compatible server. Install the optional dependencies and download a model first:
pip install transformers accelerate bitsandbytes
huggingface-cli download microsoft/phi-2 --local-dir ~/.cache/phi-2Configure the environment to use the local backend:
export LLM_BACKEND=local
export LOCAL_MODEL_PATH=~/.cache/phi-2 # or set LOCAL_MODEL_NAME
# LOCAL_SERVER_URL=http://localhost:8000 # optional OpenAI-style serverRun the evolution loop with the local model:
python scripts/run_example.py --experiment my_exp| Layer | Highlights |
|---|---|
| Data | Zero‑setup loader for any Papers‑With‑Backtest dataset (pwb_toolbox) + caching to Feather |
| Strategies | Seed templates with EVOLVE‑BLOCK markers that the LLM mutates |
| Evaluator | Deterministic Backtrader walk‑forward, JSON KPIs (Sharpe, CAGR, Calmar, DD) |
| LLM Engine | OpenAI o3 structured‑output chat → JSON diff/patch system |
| Evolution | Async controller, SQLite hall‑of‑fame, optional MAP‑Elites niches |
| Dashboard | (optional) Streamlit live view of metrics & equity curves |
alphaevolve/
├── engine.py # convenience wrapper to run the evolution loop
├── evaluator/ # data loading, metrics & Backtrader evaluation
├── evolution/ # controller, patching, islands
├── llm_engine/ # prompt builder + OpenAI client
├── strategies/ # seed strategies (EVOLVE‑BLOCK markers)
└── store/ # SQLite persistence
scripts/ # CLI entry‑points
The PromptGenome dataclass allows the LLM instructions themselves to be
evolved using a genetic algorithm. Set ENABLE_PROMPT_EVOLUTION = True in
examples/settings.py to try this feature. New prompts are mutated, evaluated
for a few iterations and stored in a separate SQLite database.
- Fork the repo & create your feature branch (
git checkout -b feat/new-feature). - Commit your changes (
git commit -m 'feat: add something'). - Push to the branch (
git push origin feat/new-feature). - Open a Pull Request.
Please run black + ruff before submitting.
MIT © 2025 Contributors
