torchcast

torchcast is a Python package for forecasting with state-space models built on top of PyTorch <http://pytorch.org>_. Its focus is training and forecasting with batches of time-series, rather than training separate models for one time-series at a time. In addition, it provides robust support for multivariate time-series, where multiple correlated measures are being forecasted.

Currently the focus of torchcast is building models that are hybrids: they are classic state-space models with the twist that every part of these models is differentiable and can take advantage of PyTorch's flexibility. For example <https://torchcast.readthedocs.io/en/latest/examples/electricity.html#Training-our-Hybrid-Forecasting-Model>_, we can use any PyTorch Module to predict the variance of forecasts or even to generate the underlying states themselves.

This repository is the work of Strong Analytics <https://www.strong.io/>_.

.. image:: docs/examples_air_quality_6_2.png

Getting Started

torchcast can be installed with pip:

.. code-block:: bash

pip install git+https://github.com/strongio/torchcast.git#egg=torchcast

torchcast requires Python >= 3.8 and PyTorch >= 1.8.

See the Quick Start <https://torchcast.readthedocs.io/en/latest/quick_start.html>_ for a simple example that will get you up to speed, or delve into the examples <https://torchcast.readthedocs.io/en/latest/examples/examples.html>_ or the API <https://torchcast.readthedocs.io/en/latest/api/api.html>_.