EngressionTS: Probabilistic time-series forecasting via Engression

Anusha Tomar, Rajdeep Pathak, and Tanujit Chakraborty

Overview

EngressionTS is a Python package for probabilistic time-series forecasting that combines neural forecasting architectures with Engression. It provides a common framework for generating probabilistic forecasts and capturing uncertainty directly through the model.

Key Features

• Probabilistic Time-Series Forecasting : Generates multiple stochastic future trajectories to model the predictive distribution rather than producing only a single point forecast.

• Model-Intrinsic Uncertainty Quantification : Introduces stochasticity through noise injection into historical inputs, allowing uncertainty to be captured directly during forecasting.

• Energy-Based Training : Uses the Energy Score as the training objective to encourage forecasts that are both close to the observed values and appropriately diverse.

• Models : Integrates deep time series forecasting architectures, including RNNs, TCNs, Transformers, NHITS, TiDE, and TSMixer, with Engression-based probabilistic forecasting.

• Unified Probabilistic Framework : Provides a common interface for training, inference, and evaluation of Engression-augmented forecasting models across different time-series datasets and architectures.

Installation

Option 1: Install engressionts from PyPI using pip:

pip install engressionts

Option 2: Install from source

git clone https://github.com/anushatomar13/engressionts.git
cd engressionts
pip install -e .

Documentation & Tutorials

• Official Documentation Website: [To be updated]
• Tutorial Notebook: Solar Energy Example Notebook (GitHub)

Citation

If you use engressionts in your research, please cite our paper:

[To be updated]

License

This project is licensed under the MIT License - see the LICENSE file for details.