Version 0.1.3

Bugfixes

• Resolved an issue where missing attributes in Gloria’s serialization module caused Gloria.to_dict(), Gloria.from_dict(), Gloria.to_json(), and Gloria.from_json() to fail (#88).

Stan Backend Improvements

Several updates have been applied across all Stan models to improve robustness, stability, and maintainability of the Stan backend. These changes significantly reduce fallback to Newton optimization, improve fit reproducibility and mitigate overshooting trend functions. All improvements are related to issue #84:

• Added normalization of the time vector to reduce sensitivity to small changes in trend rates (#95)

• Moved regressor normalization from models.py to Stan backend and changed it to Z-score normalization (#96)

• The initial parameter calculation in models.py was changed to work on the new time and response scales.

• Updated the prior for trend rate change coefficients \(\delta\) from a Lasso-like to an elastic-net–like formulation. The additional Gaussian component reduces large, abrupt trend shifts (#97)

• Removed bounds on the regressor coefficient parameter \(\beta\) to prevent the optimizer from getting stuck at parameter limits (#98)

• Consolidated shared Stan code into separate .stan files and included them via #include to simplify maintenance (#99)

• Moved prior scale calculations from the model block to the transformed data block for improved performance (#100)

Documentation

• Added Migrate from Prophet tutorial highlighting key differences and similarities between Prophet and Gloria (#46)

• Added Predictions tutorial explaining the structure and interpretation of predict() output (#45).

• Added Model Serialization tutorial on how to store and reuse fitted Gloria models (#50).

• Added Configuration Files tutorial demonstrating how to configure Gloria models and methods using TOML files (#51).

Contributors

• Benjamin Kambs

• Patrik Wollgarten