trade-learn: A Rust-Powered Quantitative Trading Backtesting Framework with 110x Performance Boost

trade-learn is a Python+Rust hybrid architecture quantitative trading backtesting framework developed by MuuYesen. It was open-sourced on GitHub on June 1, 2026 (link: https://github.com/MuuYesen/trade-learn). This framework maintains 100% semantic alignment with Backtrader, achieves over 110x performance improvement in multi-asset backtesting via a Rust-native backtesting core, and integrates a built-in causal inference mechanism to resolve spurious correlation issues in machine learning strategies, providing an efficient and robust solution for quantitative research and investment.

In the field of quantitative trading, researchers have long faced a dilemma: choosing a feature-rich but slow framework (e.g., Backtrader, which takes hours to process large-scale multi-asset backtesting) or a high-performance but lightweight solution that sacrifices semantic accuracy. trade-learn was created to resolve this contradiction. With a hybrid architecture of Python for strategy logic expression and Rust for native backtesting core, it achieves a performance leap while maintaining semantic alignment, thus improving strategy iteration efficiency.

trade-learn adopts a 'Dual-Mode & Dual-Core' architecture:

• Engine Mode: Deeply aligned with Backtrader semantics, supports core components like Analyzer, Sizer, Signal, suitable for building production-grade complex systems;
• Lite Mode: Follows the minimalism of backtesting.py, provides a lightweight Pythonic interface, supports direct model weight connection (target_weights interface), suitable for high-frequency iteration and prototype verification. Both modes share a Rust-powered Runtime to ensure consistent underlying execution efficiency.

The underlying matching engine and core computing of trade-learn are implemented in Rust, leading to significant performance improvements:

• Single-asset backtesting: 28x speedup;
• Multi-asset rebalancing scenario: Over 110x speedup. In addition, the framework is equipped with an intelligent Runner scheduling system that automatically selects processing modes, and optimizes memory layout for index enhancement scenarios, allowing developers to focus on strategy logic.

trade-learn integrates a built-in causal inference mechanism to address spurious correlation traps in machine learning strategies:

• Through the CausalSelector component, it integrates PC algorithm and FCI algorithm to strip out spurious correlation factors and retain features that directly drive returns;
• Advantages: Resists out-of-sample decay, improves strategy interpretability, seamlessly integrates with the causal-learn ecosystem, and lowers the threshold for applying academic algorithms to practical use cases.

trade-learn provides a complete research pipeline: Research (feature generation and data splitting) → Pipeline (preprocessing, model scoring, weight generation) → Portfolio (backtesting execution) → Live-style (real-time inference). It also integrates JupyterLab and MLflow for reproducible tracking. In terms of ecosystem compatibility, it supports TDX/TradingView indicator standards and is compatible with libraries like TA-Lib; deployment is flexible with minimal core dependencies, allowing on-demand expansion of a complete research environment.

Applicable Scenarios: Agile developers (fast verification via Lite mode), index enhancement managers (large-scale backtesting completed in seconds), ML researchers (full-link automation), Backtrader users (semantic alignment + high performance), causal inference explorers (spurious correlation elimination). Summary: trade-learn balances 'speed' and 'accuracy' through technological innovation, driving quantitative research from statistical correlation to causal mechanism transformation. Recommendations: The project uses the Apache 2.0 license; you can quickly try it via pip install trade-learn.