orbit

Introducing Orbit: A Python Package for Bayesian Forecasting

Orbit is an innovative Python package tailored for Bayesian time series forecasting and inference. Designed with ease of use in mind, it offers a straightforward initialize-fit-predict interface that simplifies handling time series tasks. It combines this simplicity with the robustness of probabilistic programming under the hood, making it a powerful tool for anyone dealing with time series data.

Key Features of Orbit

Orbit excels in providing various models and methods that fit different forecasting needs. Here's what it offers:

• Models Available:

  • Exponential Smoothing (ETS): Ideal for capturing seasonality and trends in data.
  • Local Global Trend (LGT): Suitable for datasets that exhibit both local fluctuations and overarching global trends.
  • Damped Local Trend (DLT): Perfect for data where trends are present but diminishing over time.
  • Kernel Time-based Regression (KTR): Best for complex datasets requiring regression analysis with time-based kernels.

• Sampling and Optimization Methods:

  • Markov-Chain Monte Carlo (MCMC): A comprehensive method for full sampling.
  • Maximum a Posteriori (MAP): Provides a single point estimate which is particularly useful when computational efficiency is necessary.
  • Variational Inference (VI): A hybrid method combining sampling with approximate distribution techniques.

Installation Options

Installing Orbit is a breeze, with multiple methods to suit your setup:

• Using pip:

  • Install directly from PyPI: pip install orbit-ml
  • Install from the source:

    git clone https://github.com/uber/orbit.git
    cd orbit
    pip install -r requirements.txt
    pip install .
    

• Using conda:

  • Via conda-forge channel: conda install -c conda-forge orbit-ml

For those interested in experimenting with the latest developments, the project can also be installed from the dev branch using pip.

Getting Started with Damped Local Trend (DLT) Model

Orbit simplifies the application of Bayesian methods with a user-friendly approach. A quick demonstration using the DLT model includes predicting time series data. After preparing the data, users can fit the model and make predictions using a few lines of code, complementing the analysis with effective visualization tools.

Community and Support

Orbit is a community-friendly project, welcoming contributors to enhance and expand the tool further. It operates under a flexible framework, which includes a code of conduct and contribution guidelines to maintain an inclusive and productive environment.

Learning and References

To understand Orbit better, users can explore comprehensive documentation and various tutorials provided. There is also rich academic backing with citations and references, making it a reliable choice for research and practical applications in forecasting.

Conclusion

Orbit is not just another forecasting tool but a robust, flexible, and user-oriented package that brings the sophistication of Bayesian methods to everyone from novices to experts in time series analysis. Whether you need basic forecasting or complex model estimations, Orbit is designed to meet diverse needs, making it a must-have in the toolkit of anyone working with time series data.