This is a boilerplate repo for a machine-learning project involving Time Series forecasting.

In particular

• It provides a cross-validation framework to ensure model are tested thoroughly and without data leakage
• It is agnostic of the type of model involved
• It is well suited for short research projects, typical of few-weeks coding bootcamps such as Le Wagon DataScience

Detailed package workflow

Architecture

• ts_boilerplate package
  • main.py comprises the main routes to be called from the CLI (train and cross-validate)
  • params.py contains project-level global variable to be set manually

• data folder contains
  • raw and clean folder should contain 2D arrays data with (axis 0) representing timesteps, and (axis 1) columns containing tagets and covariates, as per picture

    data.shape = (length, n_targets+n_covariates)

  • Xy may persist your tuple (X,y) of 3D arrays to be fed to your models

    X.shape = (n_samples, input_length, n_covariates)
    y.shape = (n_samples, output_length, n_targets)

• notebooks
  • tutorial_ts_forecasting.ipynb is a recommended read before diving into this project. It contains visuals that will help you fill global project params and understand naming conventions
  • create_dummy_tests.ipynb will help you understand how tests have been built

• tests folder detailed below

How to test your code?

Run this in your terminal from the root project folder to check your code

• pytest
• pytest -m "not optional" to only check mandatory tests
• pytest -m "not optional" -m "not slow" to also avoid tests that may be slow (involving fitting your model)

These tests require ts_boilerplate/params.py to be filled corresponding to your true project speficities

TODO

V1

• Basic training & cross-val routes, with test
• Make model fit well for univariate, multivariate (n_tagets >1) & sequences (output_sequence_lenght>1)
• Make tests pass for stride > 1
• Add tests about the model (shape of prediction, etc)

V2

• Add backtesting as main route. Very important concept to teach to students
• Refacto model.py
  • Rename pipeline.py because it may comprises the pre-processing such as scaling etc...
  • Turn into a class TsPipeline() instead of pure functions

V3

• Add tests for future-covariates
• create Makefile
  • Include DAG of the project
• publish to lewagon community