Part 2 Foundation models developed for forecasting
Now that we understand the fundamental concepts of foundation forecasting models, we can start our exploration of the major contributions. Specifically, we start with the models built exclusively for forecasting.
In chapter 3, we start with TimeGPT, which is one of the first foundation models proposed. We learn to forecast, fine-tune it, work with exogenous features, and even perform anomaly detection.
In chapter 4, we move on to Lag-Llama, which is especially suited to research purposes. We discover its architecture and pretraining protocol and learn to use it for zero-shot inference. We also study how its parameters affect its performance and how to fine-tune it.
Chapter 5 explores Chronos, a framework that adapts large language models (LLMs) to time-series forecasting. We use this univariate model for forecasting and anomaly detection, and we also fine-tune it.
In chapter 6, we study Moirai, which released one of the largest publicly available time-series datasets to further the research in the field. We learn to use Moirai with exogenous features and for anomaly detection.
This part concludes with chapter 7, where we discover TimesFM. This model is a deterministic model, which is ideal for achieving reproducible results at all times.