How to Forecast with Class Probabilities¶

This guide shows you how to fit a class-probability forecaster, obtain per-class probability predictions, and evaluate them with classification metrics. Reach for this workflow when you need per-class probabilities (for risk-aware decisions, calibration analysis, or proper scoring rules) rather than the hard labels a point forecaster predicts for categorical targets.

Prerequisites¶

• Familiarity with the fit-predict workflow (Getting Started)
• Familiarity with train/test evaluation (Evaluate Forecast Accuracy)

Try it interactively¶

Prepare Data and Train/Test Split¶

Use one of the built-in classification datasets, or prepare your own DataFrame with a "time" column and one or more string-valued target columns. Split the data before fitting so the evaluation later reflects true out-of-sample performance:

from yohou.datasets import fetch_air_quality_classification
from yohou.model_selection import train_test_split

data = fetch_air_quality_classification()
y = data.y
# DataFrame with "time" and "air_quality" columns
# air_quality values: "good", "moderate", "unhealthy", "hazardous"

y_train, y_test = train_test_split(y, test_size=24)

Fit a Class-Probability Forecaster¶

ClassProbaReductionForecaster wraps any scikit-learn classifier that supports predict_proba(). The default estimator is LogisticRegression; any classifier with fit(), predict(), and predict_proba() works:

from sklearn.ensemble import GradientBoostingClassifier
from yohou.class_proba import ClassProbaReductionForecaster

forecaster = ClassProbaReductionForecaster(
    estimator=GradientBoostingClassifier(n_estimators=100),
)
forecaster.fit(y_train, forecasting_horizon=24)

Get Predictions¶

Soft probabilities (recommended for decision-making):

y_proba = forecaster.predict_class_proba()
# Columns: vintage_time, time, air_quality_proba_good, air_quality_proba_moderate, ...

Hard labels (argmax of probabilities):

y_pred = forecaster.predict()
# Columns: vintage_time, time, air_quality

Evaluate with Classification Metrics¶

Proper scoring rules give reliable model comparisons because they reward calibrated probabilities. Prefer LogLoss or BrierScore over Accuracy for model selection:

from yohou.metrics import LogLoss, BrierScore, Accuracy

log_loss = LogLoss().fit(y_train).score(y_test, y_proba)
brier = BrierScore().fit(y_train).score(y_test, y_proba)
accuracy = Accuracy().fit(y_train).score(y_test, y_pred)

Visualize Results¶

plot_forecast auto-detects categorical and probability columns:

from yohou.plotting import plot_forecast

# Hard labels: step chart
plot_forecast(y_test, y_pred)

# Probabilities: stacked area chart
plot_forecast(y_test, y_proba)

Use plot_calibration to assess whether predicted probabilities match observed frequencies:

from yohou.plotting import plot_calibration

plot_calibration(y_proba, y_test)

See Also¶

• Class-Probability Forecasting for theory and mathematical details
• Evaluate Forecast Accuracy for the complete metrics guide
• API Reference: yohou.class_proba