v0.3.0

Computational Efficiency Improvements

• New aggregate function that generates the hierarchical time series and the aggregation constraints matrix. Improve from $O((N_{a}+N_{b})^{2}log(N_{a}+N_{b}))$ to $O((N_{a}+N_{b})$.
• Vectorization of the creation of probabilistic prediction levels, before done in for loops now performed in a single vectorized numpy call.

Evaluation Utilities

• Added scaled continuous ranked probability scores (sCRPS).
• Added mean scaled squared errors (MSSE).
• Added energy score metric.
• Added random sampling outputs to probabilistic reconcilers.
• Added core.bootstrap_reconcile method to apply over different random seeds the reconcilers and generate standard deviations.

Refactorization of the HierarchicalForecast classes

• Overall improvement of the core.reconciliation method.
• Decoupled the probabilistic reconciler classes from the mean reconciler classes.
• Decoupled fit protections from reconciliation.
• Reconciler's inputs now mostly receive mostly numpy arrays.
• Simplified and deprecated dependencies.

Documentation Improvements

• Installation guide.
• New introduction tutorial with minimal, intuitive example.
• Tutorial on evaluation of reconciliation probabilistic reconciliation baselines.

New Collaborators and HierarchicalForecast Paper

• We started a fruitful collaboration with Souhaib Ben Taieb
  and Shanika Wickramasuriya.
• We submitted the HierarchicalForecast library paper to the Journal of Machine Learning Research.

What's Changed

• [FEAT] Ignore jupyter notebooks as part of languages in #120
• [FEAT] Factorizing reverse_sigmah from HierarchicalReconciliation in #121
• [FEAT] Decoupling _reconcile, from _get_PW_matrices. in #123
• [FEAT] PW initialization in #124
• Prob Reconciler's tests location in #125
• Core Refactorization + Reconcilers.fit in #128
• CircleCI in #129
• Shared HReconciler + predict method in #131
• [FEAT] Reconciler's sample method in #133
• [FEAT] CRPS, MSSE and Energy Score metrics in #134
• time tracking utils in #135
• [FEAT] Faster creation of ProbReconciler's ordered levels in #137
• [FIX] Matplotlib and numba errors in #142
• [FIX] Circle ci integration in #141
• [BUG] PERMBU unique_id order and num_samples in #143
• [Bug] Fixed S_df categorical index ordering in #145
• [FEAT] seed/num_samples usage possibility + MSSE evaluation example in #147
• [FEAT] Faster aggregate function + Gaussian Log Score in #150
• [FIX] Documentation + Update bib reference in #156
• light improvements to readme in #157
• [FIX] Use micromamba instead of miniconda (CI) in #167
• [BUG] Added level domain protection for normality and permbu methods in #166
• Level domain protection in #169
• Omit expensive linear algebra when not necessary in MinTrace in #171
• [FIX] Add correct github link in #173
• [DOCS] Improved index, intro, quick start, and geographical forecasts in #175

New Contributors

• @melopeo made their first contribution in #157
• @mcsqr made their first contribution in #171
• @cchallu made their first contribution in #175

Full Changelog: v0.2.1...v0.3.0