Expanding-Window LCE Cross-Validation

Time-series style cross-validation for TaskLCE. Splits are made by whole batches (column with role feature, i.e. the batch_nr column of the task). Each fold trains on all batches from the very first up to a moving training end and tests on the next horizon batches; the training window therefore always starts at the beginning and expands by step_size batches between folds.

Creates a new expanding-window LCE resampling.

Fields

iters

(integer(1))
Number of resampling iterations. Only meaningful after $instantiate().

Parameters

• horizon :: integer(1)
  Number of consecutive batches used as test set in every fold. Initialized to 1.

• step_size :: integer(1)
  Number of batches between the train-end indices of consecutive folds. Initialized to 1.

• min_train_batches :: integer(1)
  Number of batches in the training set of the first fold. Required, with no default: results are sensitive to it and there is no good universal value. Must be at least the wrapped learner's minimum training requirement, which is larger for some learners (e.g. LearnerLCESplineMonotone needs five batches, LearnerLCEConformal needs n_calibration_batches + 1).

• folds :: integer(1) | NULL
  Maximum number of folds. When unset, all feasible folds are generated.