Explanation

Conceptual background for understanding how Yohou works and why it is designed the way it is. These pages complement the tutorials and how-to guides by giving you the mental models needed to use Yohou confidently on unfamiliar problems.

Foundations

• Core Concepts: The fit/observe/predict lifecycle, data formats, and the reduction approach to forecasting.
• Time Series Patterns: Trend, seasonality, cycles, and noise, and how to recognise them before choosing a method.
• Reduction Forecasting: How Yohou converts time series into supervised learning problems, and what that means for feature construction and prediction.

Data Shaping

• Panel Data: The {entity}__{variable} naming convention, the three panel strategies, and panel-aware behavior in forecasters and scorers.
• Exogenous Features: The three exogenous types (X_actual, X_future, X_forecast), step-indexed columns, and vintage alignment.
• Preprocessing: Stateful vs. stateless transformers, the BaseTransformer contract, and incremental observation in pipelines.
• Stationarity: Why non-stationary series are problematic for regression models, and how differencing and decomposition help.
• Feature Pipelines: FeaturePipeline, FeatureUnion, and ColumnTransformer: how to compose transformers and how observation_horizon propagates.

Forecasting

• Forecaster Composition: DecompositionPipeline, ColumnForecaster, ForecastedFeatureForecaster, LocalPanelForecaster, and state propagation through composite forecasters.
• Interval Forecasting: Prediction intervals, conformal coverage, and when to use SplitConformalForecaster vs. quantile regression.
• Class-Probability Forecasting: Categorical time series, calibration, and the class-probability forecaster API.
• Ensemble Forecasting: Voting strategies, error diversity, and when ensembles outperform single models.

Evaluation

• Forecast Accuracy: MAE, MASE, CRPS, and other metrics: which rewards what behavior, and common pitfalls.
• Model Selection: Why standard cross-validation fails for time series, and how expanding and sliding windows preserve temporal order.
• Weighting: Time weighting, vintage weighting, step weighting, and how they apply at both fit time and score time.
• Residual Diagnostics: How to interpret residual plots and what patterns signal unmodelled structure.
• Visualization: The plotting module, interactive Plotly figures, and choosing the right plot for your task.

Architecture

• Metadata Routing: How metadata like time_weight and vintage_weight flows through pipelines, search objects, and composite forecasters.
• Extending Yohou: Abstract base classes, parameter constraints, integration packages, and the systematic test suites for custom components.

Reference

• Glossary: Definitions of key terms used across Yohou documentation.