A Matrix-Based Machine Learning Runtime for Pine Script

Summary

This document describes a Pine Script library for building and running machine learning models. Its primitives include tensors and matrix operations, differentiable computation graphs, neural network layers, recurrent and attention operations, and sequential model helpers. It also provides training components such as optimizers, losses, metrics, learning-rate schedules, data scaling, rolling datasets, validation-related controls, and early stopping.

The library includes losses and metrics relevant to forecasting, including directional accuracy, quantile loss, multi-horizon weighted error, and a Sharpe-oriented loss. The source excerpt demonstrates the available interfaces and loss calculations, but it presents no trading experiment, model comparison, or performance evidence. It is therefore a software toolkit description rather than a validated trading method. Researchers would need to assess implementation behavior, data handling, leakage risk, and out-of-sample performance for their own models.

Key ideas

• The library supports tensor and matrix operations alongside differentiable graph execution.
• It includes common neural network components, including convolution, attention, and recurrent scans.
• Training utilities cover optimizers, loss functions, metrics, schedules, and early stopping.
• Dataset and scaling tools support windowed or rolling data workflows.
• The document provides implementation features but no evidence that models built with the library are profitable.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.