Time series forecasting using ML techniques

This repository contains the work for my MEng mini-thesis. The project investigates feature engineering techniques for time series forecasting using machine learning.

Description repo

Description of folders:

• misc code: old code, shared code or just useful snippets found during research.
• old playground notebooks: messy jupyter notebooks used to trial new ideas or first attempt to develop various models and techniques.
• pics: a place to save result figures.
• results: a place to save forecasting results.
• test data: example data used for the project

Description of files:

• one_dimensional_time_series_forecasting.py: A python module containing a class which encapsulate the whole time series forecasting pipeline. Time series to supervised ml conversion, testing-training splits, training and tuning multiple ML and DL models, evaluating performance and plotting results. There are also a few miscellanious but usefull helper functions.
• basic_feature_engineering.ipynb: A notebook which investigates:
  • normalisation
  • differencing
  • log-differencing
• spectral_feature_engineering.ipynb: A notebook which investigates spectral methods used to denoise time series signals.

Using the module

To use this work you will be primarily interested in the generalised forecasting pipeline implemented in the one_dimensional_time_series_forecasting.py module.

All dependencies required are contained within:

• requirements.txt

And can be installed using pip:

pip install -r requirements.txt