ProFiT: Program Search for Financial Trading

This repository contains a reproduction of the ProFiT framework as described in the research paper "ProFiT: Program Search for Financial Trading".
ProFiT is an LLM-driven evolutionary search algorithm that autonomously discovers and improves algorithmic trading strategies. It uses a closed feedback loop where an LLM analyzes strategy performance on backtests and proposes code-level improvements.

Key Features

• LLM-Driven Evolution: Uses Large Language Models (e.g., GPT-4) to act as both "Analyst" and "Developer" to mutate trading strategies.
• Code-Level Search: Strategies are represented as executable Python code, not abstract parameters.
• Self-Repair: Includes a repair loop where the LLM fixes syntax or runtime errors based on tracebacks.
• Walk-Forward Validation: Implements the specific 5-fold time-series split described in the paper.
• Robust Backtesting: Uses backtesting.py (via lucit-backtesting) with realistic constraints ( 0.2% commission, exclusive orders).

Installation

This project is managed with uv.

1. Install dependencies:

   uv sync

2. Set up Environment: If using the OpenAI client, set your API key:

   export OPENAI_API_KEY="sk-..."

Configuration

Edit config.yaml to control the experiment parameters:

data:
  path: "tests/data/Bitcoin_historical_data_coinmarketcap.csv"
ga:
  generations: 15  # Number of evolutionary steps
llm:
  client: "openai" # "openai" or "mock"
  model: "gpt-4o"

Running the Reproduction

To run the standard experiment (Seed: Bollinger, Fold: 0):

python main.py

Results will be saved to runs/reproduction_run, including the generated strategy code and fitness plots.

Project Structure

• profit/optimizer.py: The core ProFiT evolutionary loop (Algorithm 1).
• profit/llm/client.py: LLM interaction logic (Analysis, Improvement, Repair prompts).
• profit/seed_strategies.py: The 5 initial seed strategies and baselines.
• profit/backtester/lucit_adapter.py: Adapter for the backtesting engine.
• profit/strategy.py: Definition of the code-based Chromosome.

Differences from Original Paper

• This implementation currently runs on a single Fold per execution (configurable in main.py).
• Default config uses a Mock LLM for safety; switch to "openai" for real reproduction.

License

MIT