Vibe Coding DiscoverAgentQuant
View on GitHubAutonomous quantitative trading research platform with self-improving AI agents using adaptive harness evolution, transforms stock lists into fully backtested strategies without coding
AgentQuant is a Python research platform where a ReAct LLM agent proposes trading strategies, backtests them with walk-forward validation, reflects, and stores results in cross-session memory. It evolves its own search harness across epochs under quality gates; the zero-key demo runs offline.
Use Cases
Generate and backtest stock trading strategies from a ticker list without codingRun an analyze-hypothesize-backtest-reflect ReAct loop for strategy searchWalk-forward validation with train/validation/test splits to measure overfittingPersist cross-session SQLite memory of strategies that worked per market regimeLLM-guided parameter proposal using Tavily web search and academic research contextEvolve the agent's own research harness across epochs under quality gatesDetect look-ahead bias and data leakage before trusting benchmark resultsVisualize backtest metrics and agent memory in a Streamlit dashboardCompare grid/random/GA/DE search baselines on reproducible benchmarks
Built With
- Language
- Python
- Frameworks
- LangChain · LangGraph · Streamlit · vectorbt · yfinance · pandas · numpy · scikit-learn · hmmlearn · Pydantic · Plotly · pytest
Tags
quantitative-trading · agentic-ai · self-improving · react-agent · backtesting · harness-evolution · walk-forward-validation · portfolio-optimization · regime-detection · long-term-memory · fintech · quant-research · no-code · evaluation-harness