Time-Series CrossValidation for NN

https://github.com/DidierRLopes/timeseries-cv

2 years ago, Filipe Ramos my previous maths and probability teacher, knowing that I had a special interest in Data Science, challenged me to help him in his PhD thesis “Data Science na Modelação e Previsão de Séries Económico-financeiras: das Metodologias Clássicas ao Deep Learning”.

Although we have been discussing theory, analysis and results, my main contribution was to write the Python code behind the thesis.

As a result, I have written a python module that splits a given univariate time-series based on cross-validation techniques so that these can be fed to a Deep Neural Network (DNN) to extract training/validation/test errors.

I know that there are examples of these online, but this was made from scratch so that we could personalise it according to the thesis’ needs, and grasp better what was at stake when performing different cross-validation techniques.

The idea is given a training dataset, the package will split it into Train, Validation and Test sets, by means of either Forward Chaining, K-Fold or Group K-Fold.

As parameters the user can not only select the number of inputs (n_steps_input) and outputs (n_steps_forecast), but also the number of samples (n_steps_jump) to jump in the data to train.

The best way to install the package is as follows: pip install timeseries-cv and then use it with import tsxv. See the module developed here.