I just beat AI traders with math

(pypi.org)

5 points | by thattechgeek 12 hours ago ago

3 comments

  • thattechgeek 12 hours ago

    Hi HN,

    I built Sagan Trade after seeing every quant fund struggle with the same problem: their models work until they catastrophically fail, and no one can explain why.

    We just published research showing symbolic regression statistically significantly outperforms deep learning for trading:

    - Symbolic engine: +11.84% return, 2.46 Sharpe, -3.09% max drawdown - TFT-PINN model: -37.52% return, -2.47 Sharpe, -35.10% max drawdown - Statistical significance: p=0.0206

    The key insight: instead of black-box neural networks, we fit price/volume data to transparent polynomial+Fourier basis functions. Every signal is a mathematical equation you can inspect and audit.

    Technical highlights: - R² > 0.94 fitting accuracy on price series - FunctionGemma LLM for post-prediction explainability - Sub-50ms signal generation with Numba JIT - Three execution modes: Coordinated, Market Neutral, Long-Only

    Try it: pip install sagan-trade

    I'm here to answer questions about the methodology, results, or implementation!

    • sillysaurusx 12 hours ago

      Do you have a link to more of your symbolic engine research? I'd be curious to read about it. Particularly which polynomials and basis functions resulted in the +11.84% return. That seems hard to replicate, otherwise traders will take advantage of it right away.

    • nikolay 10 hours ago

      Why did you remove the source code?