Splits training and test data chronologically by similarities based on a K-means cluster. Can be helpful to split data chronologically keeping relevant portions in both train and test datasets. Actions: Trains cluster from input dataset. Processes batch centroid over the same dataset. For each cluster label, creates specific dataset and splits train-test chronologically with given proportions. Merges all train and test portions separately providing a final train and test database. Requirements: The original dataset needs to be ordered chronologically.
This script contains a modified version of the Cross-Validation technique that is able to deal with
time dependent data.
The i.i.d. (identically distributed independence) assumption does
not hold well to time dependent data. There is an implicit dependence
on previous observations, so the value of the objective field at a
given time is related to the values at previous (and future) times.
Traditional Cross-Validation, which performs random splits, ignores
this, resulting in serious underestimation of predictive error.
Time-Aware Cross-Validation techniques, also known as Block
Cross-Validation, divide the dataset into train and validation folds
ensuring that the validation set is always ahead of the training one.
This script offers two kinds of Time-Aware Cross Validation
techniques: Forward Chaining and Fixed Window:
Forward Chaining uses an expanding window meaning that, on each iteration, a new block of samples will be added to the
training data.
Fixed Window uses a window with fixed size that slides over all the samples.
