Find the workshop material on GitHub: https://github.com/unit8co/amld2022-forecasting-and-metalearning

Link to presentation: https://docs.google.com/presentation/d/1iViBCRa-sG_8KzJrqbyI_Vv08Zi9KmQ73Vkk9KohyKA/edit?usp=sharing

Wouldn’t it be nice if you could predict the future based on previous observations ? This is what time series forecasting is about. As such and coupled with the abundance of time series data (sequences of values obtained at successive times), time series forecasting has become an essential tool for businesses, people and institutions. It has applications in logistics, predictive maintenance, energy, manufacturing, agriculture, climate science, and many other domains.

Recently, new deep learning models have emerged producing state-of-the-art results in domains from sales and demand forecasting to energy. However, applying these models in practice can still be a challenge: from formatting the data, to implementing the models or obtaining forecasts that factor in external data.

In this workshop, we will share our learning and techniques to train and apply deep learning forecasting models in practice. We will also talk about meta-learning, which is a new avenue opened by these models. Meta-learning enables the discovery of generic patterns from a diverse set of distinct time series and improve generalization on new time series coming from potentially different datasets, opening the door for zero-shot forecasting and avoiding the cold start problem.

During the workshop, we will cover the following topics:

• How to train state-of-the-art deep-learning models such as N-BEATS and TFT
• How to leverage meta-learning properties to forecast previously unseen time series
• Backtesting your models and compare them with traditional statistic-based forecasting models

After introducing the concepts, you will have the chance to apply all the tools and techniques presented by competing against the other participants in a time series forecasting competition.

Plan

1. Intro to time series forecasting (~15 minutes)
2. Short intro to Darts (~15 minutes)
3. Hands-on session - classical forecasting with Darts (~1h15)
4. — Coffee Break — (15 minutes)
5. Introduction to machine learning for forecasting (~30 minutes)
6. Hands-on session - ML forecasting with Darts (~1h30)

Forecasting Library: Darts

Darts (https://github.com/unit8co/darts/) is an open-source Python library dedicated to make it easy to manipulate and forecast time series data. It offers an API similar to scikit-learn, with the goal of being intuitive and familiar to most, while allowing users to quickly use a wide range of forecasting models from exponential smoothing to neural networks. It will be used during the workshop as it makes many of the tasks we will present easy.

• Learn about Deep-learning forecasting methods and their applications
• Be able to evaluate model performance and compare with traditional techniques
• Be able to quickly tackle new time series forecasting problems

• Basic knowledge of Python for data science
• Basic knowledge of machine learning
• Participants need to have a way to run the provided notebooks, for instance:
  • [recommended] on a Colab environment (https://research.google.com/colaboratory/)
  • on their own laptop, with Python 3.7+ installed in a fresh virtual environment that can be used to install the required dependencies. Having an Nvidia GPU can help in some cases for training some of the deep learning models.

Find the workshop material on GitHub: https://github.com/unit8co/amld2022-forecasting-and-metalearning

Link to presentation: https://docs.google.com/presentation/d/1iViBCRa-sG_8KzJrqbyI_Vv08Zi9KmQ73Vkk9KohyKA/edit?usp=sharing