Hi everyone,
I’d like to share a new Python based open‑source time‑series forecasting framework we developed called Gloria:
- PyPI: Client Challenge
- GitHub: GitHub - e-dyn/gloria
- Documentation: Gloria documentation — Gloria 0.1.1 documentation
Gloria stems from our experience with Prophet and retains its GLM‑inspired design, but extends it with richer probabilistic modeling and finer control over model components.
From a Stan perspective, the most notable difference is distributional flexibility: where Prophet only allows normal distribution, Gloria additionally offers Poisson, Binomial, Negative Binomial, Gamma, Beta, and Beta-Binomial (with more to come). This has two main advantages:
- Count Data, proportions, or non-negative float data can be handled natively without any additional data transformations
- No need to pick additive vs. multiplicative models or linear vs logistic trends. The nonlinear link functions of the underlying distributions automatically takes care of that.
The Python library around the Stan models has a similar feel to Prophet but adds notable features, including a growing event library that goes beyond single-day “box-car” events and the ability to handle high-frequency time grids, removing Prophet’s daily-sampling quasi-limitation.
We’re actively developing Gloria and would welcome any feedback, ideas, or use cases from the Stan community.
Regards,
Benjamin