How to Integrate with Yohou¶

This guide shows you how to use Yohou-Nixtla forecasters inside Yohou's ecosystem: pipelines, hyperparameter search, scikit-learn introspection, and the observe/rewind lifecycle.

Prerequisites¶

Yohou-Nixtla installed (Getting Started)
Familiarity with Yohou's fit/predict/observe/rewind pattern

Clone a forecaster¶

All forecasters are compatible with sklearn.base.clone, producing an unfitted copy with the same hyperparameters:

from sklearn.base import clone
from yohou_nixtla import AutoARIMAForecaster

original = AutoARIMAForecaster(season_length=12, max_p=5)
cloned = clone(original)
# cloned has the same parameters but is not fitted

Cross-validation and search utilities rely on this internally.

Inspect and modify parameters¶

Use get_params and set_params to read or change parameters programmatically:

from yohou_nixtla import AutoARIMAForecaster

forecaster = AutoARIMAForecaster(season_length=12)

params = forecaster.get_params()
# {"season_length": 12, "max_p": 5, "freq": None, ...}

forecaster.set_params(season_length=4)

get_params(deep=True) includes nested transformer parameters with the <param>__<subparam> convention, so you can access, for example, target_transformer__with_mean.

Search hyperparameters with GridSearchCV¶

Use Yohou's GridSearchCV rather than sklearn's. It wraps the time series fit/predict lifecycle correctly, including the forecasting_horizon argument:

from yohou.model_selection import GridSearchCV
from yohou_nixtla import AutoARIMAForecaster

forecaster = AutoARIMAForecaster()

param_grid = {
    "season_length": [4, 12],
    "max_p": [2, 5],
    "max_q": [2, 5],
}

search = GridSearchCV(
    forecaster,
    param_grid,
    scoring="neg_mean_absolute_error",
    cv=3,
)
search.fit(y, forecasting_horizon=12)

best_forecaster = search.best_estimator_
print(search.best_params_)

Tune nested transformer parameters¶

Transformer parameters are searchable through the double-underscore convention:

from sklearn.preprocessing import StandardScaler
from yohou.model_selection import GridSearchCV
from yohou_nixtla import AutoARIMAForecaster

forecaster = AutoARIMAForecaster(target_transformer=StandardScaler())

param_grid = {
    "season_length": [12],
    "target_transformer__with_mean": [True, False],
}

search = GridSearchCV(forecaster, param_grid, cv=3)
search.fit(y, forecasting_horizon=12)

Use observe, observe_predict, and rewind¶

Yohou-Nixtla forecasters support Yohou's incremental lifecycle methods for updating or rolling back the model state as new data arrives:

from yohou_nixtla import AutoARIMAForecaster

forecaster = AutoARIMAForecaster(season_length=12)
forecaster.fit(y_train, forecasting_horizon=12)

# Observe new data without refitting
forecaster.observe(y_new)

# Observe and produce rolling predictions in one call
y_pred = forecaster.observe_predict(y_new)

# Roll back to a previous state
forecaster.rewind(steps=1)

observe ingests new observations incrementally. observe_predict combines observation with rolling predictions, advancing the window by the forecasting horizon at each stride. rewind rolls the internal state back without refitting.