How-to Guides¶

Task-oriented recipes for common Yohou workflows. Each guide assumes you have completed the tutorials and are familiar with the basics.

Setup¶

Installation: Install Yohou with pip, uv, or conda, including optional extras and development setup.

Choose a Forecasting Method: Pick the right forecaster and reduction strategy for your data.
Build Reduction Forecasters: Create lag-feature forecasters, combine feature transformers, and choose a reduction strategy.
Produce Prediction Intervals: Wrap a point forecaster with SplitConformalForecaster to generate calibrated prediction intervals.
Combine Forecasters with Ensembles: Use voting ensembles to combine multiple forecasters.
Forecast with CatBoost: Use gradient-boosted trees for point and interval time series forecasting.
Forecast with Class Probabilities: Predict discrete class labels with calibrated probabilities.

Evaluate Forecast Accuracy: Score predictions with point, interval, and classification metrics.
Evaluate with Multi-vintage Scoring: Generate forecasts from successive observation points and break down errors by vintage and horizon step.
Tune Forecaster Hyperparameters: Search and tune forecaster hyperparameters with GridSearchCV and RandomizedSearchCV.

Visualize Forecasts: Plot forecast output against actuals using plot_forecast, plot_residuals, and plot_calibration.
Visualize and Compare Model Scores: Plot per-step accuracy, per-vintage trends, score distributions, and model comparisons.

Work with Panel Data: Handle multiple related time series with the __ column naming convention.
Use Exogenous Features: Incorporate external predictors using X_actual, X_future, and X_forecast.
Work with Forecast Vintages: Prepare, align, and predict with X_forecast features from upstream models stamped with an issuance time.
Add Calendar and Time Features: Engineer temporal features like day of week, month, and holidays.
Use Time Weighting: Apply non-uniform weights to emphasize recent or seasonal observations.

Use Preprocessing Transformers: Apply ColumnTransformer, FunctionTransformer, sklearn scalers, and window transformers.
Clean and Resample Time Series: Detect type mismatches, validate value ranges, and change frequency with Downsampler and Upsampler.
Compose Feature Pipelines: Build FeaturePipeline chains, combine branches with FeatureUnion, and compose panel-aware pipelines.
Apply Stationarity Transforms: Remove trends and seasonality with SeasonalDifferencing and DecompositionPipeline.
Apply Signal Processing: Smooth noise with digital filters and inspect the frequency spectrum.

Save and Load Forecasters: Serialize fitted forecasters to disk and reload them for batch scoring or deployment.
Handle Missing Data: Detect and fill gaps in a time series before they propagate through the pipeline.
Handle Outliers: Clip or flag outliers and limit their effect on conformal prediction intervals.
Handle Complex Seasonality: Choose between Fourier terms, nested decomposition, and feature engineering for multi-period seasonality.
Handle Short Series: Fallback strategies for series too short for seasonal estimation, cross-validation, or scaled metrics.
Handle Long Series: Limit history with observation_horizon, weight recent errors, and resample to the right frequency.

Create a Point Forecaster: Implement a custom point forecaster by subclassing BasePointForecaster.
Create an Interval Forecaster: Implement a custom interval forecaster by subclassing BaseIntervalForecaster.
Create a Class-Probability Forecaster: Implement a custom categorical forecaster by subclassing BaseClassProbaForecaster.
Create a Transformer: Implement a custom time series transformer by subclassing BaseTransformer.
Create Custom Scorers: Implement a custom evaluation metric by subclassing BasePointScorer or BaseIntervalScorer.
Contribute to Yohou: Set up a development environment, run tests, and submit changes.