ForecastingGridSearchCV

Perform grid-search cross-validation to find optimal model parameters.

Quickstart

from sktime.forecasting.model_selection import ForecastingGridSearchCV

estimator = ForecastingGridSearchCV(forecaster, cv, param_grid, scoring=None, strategy='refit', update_behaviour='full_refit', refit=True, tune_by_instance=False, tune_by_variable=False, verbose=0, return_n_best_forecasters=1, error_score=nan, backend='loky', backend_params=None, n_jobs='deprecated')

Parameters(15)

forecastersktime forecaster, BaseForecaster instance or interface compatible

The forecaster to tune, must implement the sktime forecaster interface. sklearn regressors can be used, but must first be converted to forecasters via one of the reduction compositors, e.g., via make_reduction

cvsktime time series splitter

Re-sampling strategy for cross-validation, must be an instance of a sktime time series splitter, e.g. SlidingWindowSplitter()

param_griddict or list of dictionaries

Model tuning parameters of the forecaster to evaluate

Encodes the search spaces for the parameters, a countable set of parameter combinations. Parsed as following:

• if dict, then the keys are parameter names (str) and the values are lists of parameter settings to try as values. Parameters not set are kept as defaults. A single non-default parameter setting can be provided as a list with one element, e.g., {"window_length": [5]}.

• if list of dict, then each element of the list is a separate dict as described above, and the search space is the union of the search spaces defined by each dict.

scoringsktime metric (BaseMetric), str, or callable, optional (default=MAPE)

scoring metric to use in tuning the forecaster

• sktime metric objects (BaseMetric) descendants can be searched with the registry.all_estimators search utility, for instance via all_estimators("metric", as_dataframe=True)

• If callable, must have signature (y_true: 1D np.ndarray, y_pred: 1D np.ndarray) -> float, with np.ndarray being of the same length, and lower being better.

• If str, uses registry.resolve_alias to resolve to one of the above. Valid strings are valid registry.craft specs, which include string repr-s of any BaseMetric object, e.g., “MeanSquaredError()”; and keys of registry.ALIAS_DICT referring to metrics.

• If None, defaults to MeanAbsolutePercentageError()

strategy{“refit”, “update”, “no-update_params”}, optional, default=”refit”

data ingestion strategy in fitting cv, passed to evaluate internally defines the ingestion mode when the forecaster sees new data when window expands

• "refit" = a new copy of the forecaster is fitted to each training window

• "update" = forecaster is updated with training window data, in sequence provided

• "no-update_params" = fit to first training window, re-used without fit or update

update_behaviourstr, optional, default = “full_refit”

one of {“full_refit”, “inner_only”, “no_update”} behaviour of the forecaster when calling update

• "full_refit" = both tuning parameters and inner estimator refit on all data seen

• "inner_only" = tuning parameters are not re-tuned, inner estimator is updated

• "no_update" = neither tuning parameters nor inner estimator are updated

refitbool, optional (default=True)

Whether to refit the forecaster with the best parameters on the entire data.

• True = refit the forecaster with the best parameters on the entire data in fit

• False = no refitting takes place. The forecaster cannot be used to predict. This is to be used to tune the hyperparameters, and then use the estimator as a parameter estimator, e.g., via get_fitted_params or PluginParamsForecaster.

tune_by_instancebool, optional (default=False)

Whether to tune parameter by each time series instance separately, in case of Panel or Hierarchical data passed to the tuning estimator. Only applies if time series passed are Panel or Hierarchical.

• If True, clones of the forecaster will be fit to each instance separately, and are available in fields of the forecasters_ attribute. Has the same effect as applying ForecastByLevel wrapper to self.

• If False, the same best parameter is selected for all instances.

tune_by_variablebool, optional (default=False)

Whether to tune parameter by each time series variable separately, in case of multivariate data passed to the tuning estimator. Only applies if time series passed are strictly multivariate.

• If True, clones of the forecaster will be fit to each variable separately, and are available in fields of the forecasters_ attribute. Has the same effect as applying ColumnEnsembleForecaster wrapper to self.

• If False, the same best parameter is selected for all variables.

verbose: int, optional (default=0)

Verbosity level. The higher, the more messages.

return_n_best_forecastersint, default=1

In case the n best forecaster should be returned, this value can be set and the n best forecasters will be assigned to n_best_forecasters_. Set return_n_best_forecasters to -1 to return all forecasters.

return_train_scorebool, optional (default=False)

Whether to include train scores in cv_results_.

If True, the cv_results_ attribute will store training scores for each parameter combination. This can be useful for diagnosing overfitting, but may increase the memory usage of the estimator.

If False, the cv_results_ attribute will not be created.

error_score“raise” or numeric, default=np.nan

Value to assign to the score if an exception occurs in estimator fitting. If set to “raise”, the exception is raised. If a numeric value is given, FitFailedWarning is raised.

backend{“dask”, “loky”, “multiprocessing”, “threading”,”ray”}, by default “loky”.

Runs parallel evaluate if specified and strategy is set as “refit”.

• “None”: executes loop sequentally, simple list comprehension

• “loky”, “multiprocessing” and “threading”: uses joblib.Parallel loops

• “joblib”: custom and 3rd party joblib backends, e.g., spark

• “dask”: uses dask, requires dask package in environment

• “ray”: uses ray, requires ray package in environment

Recommendation: Use “dask” or “loky” for parallel evaluate. “threading” is unlikely to see speed ups due to the GIL and the serialization backend (cloudpickle) for “dask” and “loky” is generally more robust than the standard pickle library used in “multiprocessing”.

backend_paramsdict, optional

additional parameters passed to the backend as config. Directly passed to utils.parallel.parallelize. Valid keys depend on the value of backend:

• “None”: no additional parameters, backend_params is ignored

• “loky”, “multiprocessing” and “threading”: default joblib backends any valid keys for joblib.Parallel can be passed here, e.g., n_jobs, with the exception of backend which is directly controlled by backend. If n_jobs is not passed, it will default to -1, other parameters will default to joblib defaults.

• “joblib”: custom and 3rd party joblib backends, e.g., spark. any valid keys for joblib.Parallel can be passed here, e.g., n_jobs, backend must be passed as a key of backend_params in this case. If n_jobs is not passed, it will default to -1, other parameters will default to joblib defaults.

• “dask”: any valid keys for dask.compute can be passed, e.g., scheduler

• “ray”: The following keys can be passed:

  • “ray_remote_args”: dictionary of valid keys for ray.init

  • “shutdown_ray”: bool, default=True; False prevents ray from shutting

    down after parallelization.

Examples

>>> from sktime.datasets import load_shampoo_sales
>>> from sktime.forecasting.model_selection import ForecastingGridSearchCV
>>> from sktime.split import ExpandingWindowSplitter
>>> from sktime.forecasting.naive import NaiveForecaster
>>> y = load_shampoo_sales ()
>>> fh = [1, 2, 3 ]
>>> cv = ExpandingWindowSplitter (fh = fh)
>>> forecaster = NaiveForecaster ()
>>> param_grid = { "strategy": ["last", "mean", "drift" ]}
>>> gscv = ForecastingGridSearchCV (
... forecaster = forecaster,
... param_grid = param_grid,
... cv = cv)
>>> gscv. fit (y) ForecastingGridSearchCV(
... )
>>> y_pred = gscv. predict (fh) Advanced model meta-tuning (model selection) with multiple forecasters together with hyper-parametertuning at same time using sklearn notation:
>>> from sktime.datasets import load_shampoo_sales
>>> from sktime.forecasting.exp_smoothing import ExponentialSmoothing
>>> from sktime.forecasting.naive import NaiveForecaster
>>> from sktime.split import ExpandingWindowSplitter
>>> from sktime.forecasting.model_selection import ForecastingGridSearchCV
>>> from sktime.forecasting.compose import TransformedTargetForecaster
>>> from sktime.forecasting.theta import ThetaForecaster
>>> from sktime.transformations.impute import Imputer
>>> y = load_shampoo_sales ()
>>> pipe = TransformedTargetForecaster (steps = [
... ("imputer", Imputer ()),
... ("forecaster", NaiveForecaster ())])
>>> cv = ExpandingWindowSplitter (
... initial_window = 24,
... step_length = 12,
... fh = [1, 2, 3 ])
>>> gscv = ForecastingGridSearchCV (
... forecaster = pipe,
... param_grid = [{
... "forecaster": [NaiveForecaster (sp = 12)],
... "forecaster__strategy": ["drift", "last", "mean" ],
... },
... {
... "imputer__method": ["mean", "drift" ],
... "forecaster": [ThetaForecaster (sp = 12)],
... },
... {
... "imputer__method": ["mean", "median" ],
... "forecaster": [ExponentialSmoothing (sp = 12)],
... "forecaster__trend": ["add", "mul" ],
... },
... ],
... cv = cv,
... )
>>> gscv. fit (y) ForecastingGridSearchCV(
... )
>>> y_pred = gscv. predict (fh = [1, 2, 3 ])