tangyum233/crypto-strategy-research-agent

Crypto Strategy Research Agent is a small research and portfolio project that experiments with automatically improving a cryptocurrency trading strategy. It is not a live trading program and does not offer investment advice. The project works with historical 30 minute candlestick data for the ETH/USDT trading pair and a starting Backtrader strategy. The system runs a backtest, then hands the results to a small pipeline built with LangGraph, a tool for managing multi step workflows. Three roles take turns reviewing the strategy. An analyst looks at the returns, drawdown, win rate, and trade history to find problems. A designer turns those problems into concrete rules for entering trades, exiting trades, setting stop losses, and sizing positions. A coder writes those rules into a new strategy file in Python. Before a new strategy is saved, the code is checked for valid syntax and structure. The previous version is backed up before being replaced. The backtest then runs again on the new strategy. This loop repeats until the strategy reaches a performance target or the process completes a set number of rounds. The project ships with a demo mode that needs no API key and no external model. It runs the whole analyze, design, code, check, and retest loop using the strategy already in the repository, so a new user can see the process work end to end before connecting anything real. Switching to a real language model just means copying an example settings file, adding an API key, and choosing a provider such as OpenAI, with the option to use a different model for each of the three roles. Output from each run lands in a results folder: a summary of the latest backtest, a full table of trades, and a backup of each strategy version before it gets overwritten. The project is written in Python and depends on Backtrader for backtesting and LangGraph for coordinating the roles.