microsoft/qlib

Qlib is an open-source platform from Microsoft Research designed for quantitative investment research using machine learning and AI. It addresses the challenge that building an end-to-end algorithmic trading research pipeline is enormously complex: you need to fetch and clean financial data, engineer features, train predictive models, backtest strategies, and finally manage a portfolio, and each step must integrate cleanly with the others. Qlib provides a unified framework that handles the entire workflow. For data, it offers a high-performance time-series database that can store and serve daily or minute-level market data efficiently. Features (called "alpha factors" in quant research) can be defined using an expression engine and are cached and recomputed automatically. On top of clean data, Qlib ships with implementations of dozens of machine learning models including LightGBM, deep learning models like LSTM, Transformers, and graph neural networks. These models predict future stock returns, which are then used to rank securities and construct a portfolio. The backtesting engine simulates how a trading strategy would have performed historically, accounting for realistic order execution and transaction costs. A nested decision framework allows modeling high-frequency trading strategies as well as daily rebalancing. Qlib also supports reinforcement learning approaches, where an AI agent learns to trade by interacting with simulated market environments. More recently, Qlib has been integrated with a companion project called RD-Agent, which uses large language models to automate parts of the research and development process, discovering new alpha factors and optimizing models with minimal human intervention. You would use Qlib if you are a quantitative researcher, data scientist, or academic working on systematic investment strategies and want a production-tested, extensible Python toolkit rather than building everything from scratch. The stack is Python, with PyTorch for deep learning models, LightGBM for gradient boosting, and a custom binary data storage layer for high-performance data access.