No hyperparameter search space
No search will be performed: Either none of the selected algorithms' parameters has more than one permitted value, or the search space is limited to a single combination.

The settings define how the search for the best model hyperparameters is done.
During the hyperparameters optimization phase, each set of hyperparameters (defined in the Algorithms page) will be evaluated with the metric defined in the Metrics page.

Search Strategy

Strategy

Traverses all the combinations of the selected values.

Samples random combinations of hyperparameters on the selected ranges. Generally performs better than grid search with identical computing resources.

Samples random combinations of hyperparameters on the selected ranges and keeps track of past experiments to optimize hyperparameters in future experiments.

Search parameters

The search will conclude when either the time or search space limit is reached.

Search space limit

Maximum number of hyperparameter combinations to explore. Set to 0 to unconstrain.

Time limit

In minutes. Set to 0 to unconstrain.

Random state

Modify the seed to change hyperparameters sampling

Randomize grid search

Shuffle the grid before doing the search (using the random state)

Cross-validation

{{ mlTaskDesign.modeling.gridSearchParams.mode === 'TIME_SERIES_KFOLD' ? 'Each validation fold' : 'The validation set' }} contains {{ prettyTimeSteps(mlTaskDesign.evaluationParams.testSize * mlTaskDesign.timestepParams.numberOfTimeunits, mlTaskDesign.timestepParams.timeunit) }} ({{ uiState.numberOfHorizonsInTest }} forecast {{ 'horizon' | plurify: uiState.numberOfHorizonsInTest }} of {{ prettyTimeSteps(mlTaskDesign.predictionLength * mlTaskDesign.timestepParams.numberOfTimeunits, mlTaskDesign.timestepParams.timeunit) }}). {{ uiState.numberOfHorizonsInTest > 1 ? 'For each' : 'In the' }} forecast horizon, the models are evaluated on the last {{ prettyTimeSteps((mlTaskDesign.predictionLength - mlTaskDesign.evaluationParams.gapSize) * mlTaskDesign.timestepParams.numberOfTimeunits, mlTaskDesign.timestepParams.timeunit) }} (the first {{ prettyTimeSteps(mlTaskDesign.evaluationParams.gapSize * mlTaskDesign.timestepParams.numberOfTimeunits, mlTaskDesign.timestepParams.timeunit) }} (gap) {{ 'is' | plurify: mlTaskDesign.evaluationParams.gapSize * mlTaskDesign.timestepParams.numberOfTimeunits : 'are' }} skipped).

Horizons in validation {{ mlTaskDesign.modeling.gridSearchParams.mode === 'TIME_SERIES_SINGLE_SPLIT' ? 'set' : 'fold' }}

{{ prettyTimeSteps(uiState.numberOfHorizonsInTest * uiState.predictionLength * uiState.timestepParams.numberOfTimeunits, uiState.timestepParams.timeunit) }}

Cross-validation strategy

Number of folds

Fold offset

Add an offset of one forecast horizon between the hyperparameter search validation (resp. cross-test evaluation) folds

Equal duration train set folds

Create rolling train set folds of equal duration.

Cross-validation strategy

Since time-based ordering is enabled, cross-validation strategies take the time variable into account

Split ratio

Number of folds

Random state

Random state for cross-validation

Stratified

Keeps the same {{ isCausalPrediction() | targetRoleName}} distribution in the splits. See .

Grouped

Rows with same group column value are assigned to same fold. See .

Group column

Column containing the k-fold groups

                            
                                    The code must define a variable named cv that follows the scikit-learn crossvalidation object protocol