Foundation Models for Time Series

The only course you need to master large time models for time series analysis

Study and apply the latest state-of-the-art foundation models for time series data. The field is shifting towards large pretrained models, and experience with these models is a must-have for serious practitioners.

In this 5h course, we cover:

Module 1: Introduction

• Define foundation models
• Understand the transformer architecture from a time series standpoint
• Discover the benefits and drawbacks of using foundation models
• Understand the fundamental concepts of pretraining, fine-tuning and transfer learning
• Build your own tiny foundation model

Module 2: TimeGPT

TimeGPT is now mentioned in many job requirements for companies like DoorDash, Tesla and OpenAI.

• Understand TimeGPT's architecture and pretraining protocol
• Work with TimeGPT for:
  • zero-shot forecasting
  • forecasting with external variables (exogenous features)
  • explainability with SHAP values
  • fine-tuning
  • multivariate forecasting

Module 3: Chronos

An open-source model from AWS AI Labs that adapts the T5 language model for forecasting.

• Understand its architecture and pretraining protocol
• Apply Chronos for:
  • Zero-shot forecasting
  • Forecasting with exogenous features in AutoGluon
  • Fine-tuning in AutoGluon

Module 4: Chronos-2

The latest iteration of Chronos with critical capabilities like multivariate forecasting

• Understand its architecture and pretraining protocol
• Apply Chronos-2 for:
  • Zero-shot forecasting
  • Forecasting with exogenous features
  • Multivariate forecasting
  • Fine-tuning

Module 5: Moirai, Moirai-MoE, Moirai2

A family of large time models developed by Salesforce and originally the first multivariate large time model

• Understand its architecture and pretraining protocol
• Apply Moirai for:
  • Zero-shot forecasting
  • Forecasting with exogenous features
  • Multivariate forecasting

Module 6: TimesFM

Google's open-source large time model.

• Understand its architecture and pretraining protocol
• Apply TimesFM for:
  • Zero-shot forecasting
  • Forecasting with exogenous features

 Module 7: Sundial

A foundation model that introduces the TimeFlow loss function and the massive TimeBench dataset.

• Understand its architecture and pretraining protocol
• Apply Sundial for:
  • Zero-shot forecasting

Module 8: TiRex

The big comeback of the LSTM as it is the only foundation model not relying on the transformer.

• Understand its architecture and pretraining protocol
• Apply TiRex for:
  • Zero-shot forecasting

Bonus Module 1: Anomaly detection

Use large time models for anomaly detection. COMING SOON!