docs: auto-update documentation for 4.4.0

docs/source/user_guide/concepts/beam.rst

@@ -139,9 +139,10 @@ three dimensions. Each dimension answers a different operational question:
 **Configurable metrics.** Metrics are supplied as provider objects. BEAM includes
 providers such as ``RMAEProvider`` (Relative Mean Absolute Error) and ``RCRPSProvider``
 (Relative Continuous Ranked Probability Score) for probabilistic evaluation. Interval
-metrics such as ``RCSProvider`` and
-``RIQDProvider`` help assess quantile calibration and sharpness for symmetric
-quantile ranges. You can implement custom providers to add domain-specific metrics.
+metrics such as ``RCSProvider`` (Regression Coverage Score, measuring the fraction of
+observations inside predicted quantile bounds) and ``RIQDProvider`` (Relative Inter-Quantile
+Distance) help assess quantile calibration and sharpness for symmetric quantile ranges.
+You can implement custom providers to add domain-specific metrics.
 
 .. code-block:: python
 
@@ -249,4 +250,4 @@ benchmarks (hundreds of targets, multiple models) remain tractable.
    - :ref:`concept_models` for the forecasting models that BEAM evaluates.
    - :ref:`concept_metalearning` for how BEAM results inform model selection decisions.
    - :doc:`/user_guide/guides/backtesting_tutorial` for a hands-on walkthrough of setting up and running a backtest.
-   - :doc:`/api/beam` for the full openstef-beam API reference.
+   - :doc:`/api/beam` for the full openstef-beam API reference.

docs/source/user_guide/guides/probabilistic_forecasting.rst

@@ -253,15 +253,16 @@ Calibration quality can be assessed by comparing expected vs. observed quantile
 Key metrics for probabilistic forecast quality include:
 
 - **Calibration error**: the difference between expected and observed coverage per quantile
-- **Regression Coverage Score (RCS)**: the fraction of actual values inside a prediction interval such as P10-P90
+- **Regression Coverage Score (RCS)**: the fraction of actual values falling within a symmetric prediction interval (e.g., P10-P90). A calibrated 90% interval should yield an RCS close to 0.9. Available in BEAM evaluation pipelines via :class:`~openstef_beam.evaluation.metric_providers.RCSProvider`.
 - **Sharpness**: the width of prediction intervals (narrower is better, given proper calibration)
 - **Pinball loss**: the proper scoring rule for quantile forecasts, penalizing both miscalibration and lack of sharpness
 
-See :doc:`/user_guide/guides/backtesting_tutorial` for how to evaluate forecast quality on historical data.
+See :doc:`/user_guide/guides/backtesting` for how to evaluate forecast quality on historical data.
 
 .. seealso::
 
    - :doc:`/user_guide/guides/forecasting` for the overall forecasting workflow (fitting, predicting, model selection).
    - :doc:`/user_guide/concepts/models` for understanding how different model types compare.
-   - :doc:`/user_guide/guides/backtesting_tutorial` for evaluating forecast performance systematically.
+   - :doc:`/user_guide/guides/backtesting` for evaluating forecast performance systematically.
    - :doc:`/user_guide/guides/reliability_fallback` for operational concerns like fallback behavior when data is missing.
+```