How to Tune Composed Forecasters¶

This guide shows you how to tune hyperparameters in composed forecasters (pipelines and combinations that nest multiple components) using OptunaSearchCV.

Interactive notebook

See the companion notebook for a runnable example. View · Open in marimo

Prerequisites¶

Yohou-Optuna installed (Getting Started)
Familiarity with Yohou's composition API (DecompositionPipeline, ColumnForecaster)

Use Double-Underscore for Nested Parameters¶

Composed forecasters expose their components' parameters via the double-underscore convention. A PointReductionForecaster with an estimator and a feature_transformer exposes nested parameters like estimator__alpha and feature_transformer__lag:

import optuna
from optuna.distributions import FloatDistribution, IntDistribution
from sklearn.linear_model import Ridge

from yohou.metrics import MeanAbsoluteError
from yohou.model_selection import ExpandingWindowSplitter
from yohou.point import PointReductionForecaster
from yohou.preprocessing import LagTransformer
from yohou_optuna import OptunaSearchCV, Sampler

forecaster = PointReductionForecaster(
    estimator=Ridge(),
    feature_transformer=LagTransformer(lag=6),
)

param_distributions = {
    "estimator__alpha": FloatDistribution(0.001, 100.0, log=True),
    "feature_transformer__lag": IntDistribution(3, 24),
}

search = OptunaSearchCV(
    forecaster=forecaster,
    param_distributions=param_distributions,
    scoring=MeanAbsoluteError(),
    sampler=Sampler(sampler=optuna.samplers.TPESampler, seed=42),
    n_trials=20,
    cv=ExpandingWindowSplitter(n_splits=3, test_size=24),
)

search.fit(y_train, forecasting_horizon=24)

Call forecaster.get_params(deep=True) to discover the full set of tunable parameter names.

Tune a DecompositionPipeline¶

DecompositionPipeline chains multiple forecasters that each remove a component (trend, seasonality) from the series. It takes a list of (name, forecaster) tuples. Nested parameters use the step name as prefix:

from yohou.compose import DecompositionPipeline
from yohou.stationarity import PolynomialTrendForecaster, FourierSeasonalityForecaster

pipeline = DecompositionPipeline(
    forecasters=[
        ("trend", PolynomialTrendForecaster()),
        ("season", FourierSeasonalityForecaster(seasonality=12)),
    ]
)

param_distributions = {
    "trend__degree": IntDistribution(1, 3),
    "season__seasonality": IntDistribution(6, 24),
    "season__harmonics": IntDistribution(1, 6),
}

Call pipeline.get_params(deep=True) to see all available parameter paths.

Tune a ColumnForecaster¶

ColumnForecaster fits one forecaster per target column (multivariate targets). It takes a list of (name, forecaster, columns) tuples. Tune parameters for each column's forecaster using the component name as prefix:

from yohou.compose import ColumnForecaster

column_forecaster = ColumnForecaster(
    forecasters=[
        ("sales", PointReductionForecaster(estimator=Ridge()), "sales"),
        ("returns", PointReductionForecaster(estimator=Ridge()), "returns"),
    ]
)

param_distributions = {
    "sales__estimator__alpha": FloatDistribution(1e-4, 10.0, log=True),
    "returns__estimator__alpha": FloatDistribution(1e-4, 10.0, log=True),
}