Skip to content

How to Combine Forecasters with Ensembles

This guide shows you how to combine multiple forecasters into a voting ensemble for more stable predictions.

Prerequisites

Try it interactively

  • How to Combine Classification Forecasters

    Build classification ensembles with VotingClassProbaForecaster using soft and hard voting strategies.

    View · Open in marimo

  • How to Combine Forecasters with VotingPointForecaster

    Build point ensembles with VotingPointForecaster using mean, weighted, and median aggregation strategies.

    View · Open in marimo

  • How to Combine Interval Forecasters

    Build interval ensembles with VotingIntervalForecaster using envelope, mean, and median aggregation strategies.

    View · Open in marimo

1. Create a Point Ensemble

Pass named (name, forecaster) tuples to VotingPointForecaster:

from sklearn.linear_model import Ridge
from sklearn.ensemble import RandomForestRegressor
from yohou.ensemble import VotingPointForecaster
from yohou.point import PointReductionForecaster
from yohou.datasets import fetch_electricity_demand

data = fetch_electricity_demand()
y = data.frame

ridge = PointReductionForecaster(estimator=Ridge())
rf = PointReductionForecaster(estimator=RandomForestRegressor(n_estimators=50))

ensemble = VotingPointForecaster(
    forecasters=[("ridge", ridge), ("rf", rf)],
    method="mean",
)
ensemble.fit(y, forecasting_horizon=24)
y_pred = ensemble.predict()

To favor one model over another, pass weights:

weighted = VotingPointForecaster(
    forecasters=[("ridge", ridge), ("rf", rf)],
    method="mean",
    weights=[0.3, 0.7],  # favor random forest
)

Set method="median" for robustness against outlier predictions (weights are ignored with median aggregation).

2. Ensemble Interval Forecasters

VotingIntervalForecaster combines prediction intervals from multiple interval forecasters such as SplitConformalForecaster:

from yohou.ensemble import VotingIntervalForecaster
from yohou.interval import SplitConformalForecaster

interval_ensemble = VotingIntervalForecaster(
    forecasters=[
        ("conf_ridge", SplitConformalForecaster(
            point_forecaster=PointReductionForecaster(estimator=Ridge()),
        )),
        ("conf_rf", SplitConformalForecaster(
            point_forecaster=PointReductionForecaster(estimator=RandomForestRegressor()),
        )),
    ],
    method="envelope",  # most conservative: min of lowers, max of uppers
)
interval_ensemble.fit(y, forecasting_horizon=24, coverage_rates=[0.9])
y_interval = interval_ensemble.predict_interval()

Available method values: "envelope" (default, most conservative), "mean", "median".

3. Ensemble Classification Forecasters

VotingClassProbaForecaster combines class probability predictions:

from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from yohou.ensemble import VotingClassProbaForecaster
from yohou.class_proba import ClassProbaReductionForecaster
from yohou.datasets import fetch_air_quality_classification

class_data = fetch_air_quality_classification()
y_class = class_data.y

class_ensemble = VotingClassProbaForecaster(
    forecasters=[
        ("lr", ClassProbaReductionForecaster(estimator=LogisticRegression())),
        ("rf", ClassProbaReductionForecaster(estimator=RandomForestClassifier())),
    ],
    method="soft",  # weighted average of probabilities
)
class_ensemble.fit(y_class, forecasting_horizon=24)
y_proba = class_ensemble.predict_class_proba()

Use method="hard" for majority voting (argmax per base model, then mode).

4. Speed Up with Parallel Fitting

All voting forecasters accept n_jobs to fit base models in parallel:

ensemble = VotingPointForecaster(
    forecasters=[("ridge", ridge), ("rf", rf)],
    n_jobs=-1,  # use all available cores
)

5. Use with Panel Data

All voting forecasters support panel data automatically. Pass a DataFrame with __ separated panel columns, and each base forecaster receives the full panel. Aggregation happens per group:

from yohou.datasets import fetch_tourism_monthly

bunch = fetch_tourism_monthly()
y_panel = bunch.frame.select(
    ["time", "T187__tourists", "T188__tourists", "T189__tourists"]
).drop_nulls()

ridge = PointReductionForecaster(estimator=Ridge())
rf = PointReductionForecaster(estimator=RandomForestRegressor(n_estimators=50))

panel_ensemble = VotingPointForecaster(
    forecasters=[("ridge", ridge), ("rf", rf)],
)
panel_ensemble.fit(y_panel, forecasting_horizon=12)
y_pred_panel = panel_ensemble.predict()

The output contains one column per group, each with the ensemble's aggregated prediction.

See Working with Panel Data for panel data preparation and forecasting.

6. Evaluate with Rolling Predictions

Voting forecasters inherit the stateful lifecycle, so you can evaluate an ensemble walk-forward with observe_predict. It steps through the test set in batches of stride, emitting one vintage per batch without refitting:

from yohou.model_selection import train_test_split

y_train, y_test = train_test_split(y, test_size=48)

ensemble.fit(y_train, forecasting_horizon=24)
y_pred = ensemble.observe_predict(y=y_test, stride=1, forecasting_horizon=24)

The resulting multi-vintage DataFrame feeds directly into per-step and per-vintage scoring. See How to Evaluate Forecast Accuracy for the scoring workflow.

See Also