about this book
This book is meant to give you all the necessary knowledge to use large time models in the most optimal way and adapt them to your own use cases. We begin by exploring the transformer architecture, which still powers most foundation forecasting models. Then we attempt to build a tiny foundation model to experiment with concepts such as pretraining, fine-tuning, and transfer learning. This experience is a great way to appreciate the challenges of building a truly foundational model for forecasting.
Next, we explore foundation models specifically built for time-series forecasting, from TimeGPT to TimesFM. Then we experiment with LLMs applied to forecasting. We explore each method’s inner workings and pretraining procedures, which dictate the model’s capabilities and optimal use cases. That way, you’ll understand when to use a particular model and how to use it optimally. The book concludes with an experiment that draws on all the methods we explored throughout the book.