torchcast is a Python package for forecasting with state-space models built on top of PyTorch. 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, 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.
torchcast can be installed with pip:
pip install git+https://github.com/strongio/torchcast.git#egg=torchcasttorchcast requires Python >= 3.8 and PyTorch >= 1.8.
See the Quick Start for a simple example that will get you up to speed, or delve into the examples or the API.