Modern forecasting with sktime - foundation models, agents, and benchmarks

Forecasting is evolving rapidly, with deep learning, foundation models, and agentic AI opening up new possibilities while making it increasingly difficult to compare methods fairly and identify the right approach for a given application. This hands-on workshop provides a practical introduction to modern forecasting with sktime, demonstrating how a unified framework enables users to experiment with a broad ecosystem of forecasting methods while maintaining reproducibility and transparency. Rather than focusing on individual algorithms in depth, the workshop follows a use-case-driven overview of the forecasting workflow. Participants will learn how to work with classical statistical models, modern deep learning architectures, and emerging foundation models through a consistent interface, while also exploring the latest agentic capabilities for automating forecasting tasks. The emphasis is on understanding how these components fit together in real-world forecasting pipelines and how to move efficiently from experimentation to evaluation. A central theme of the workshop is trustworthy benchmarking. Participants will learn how to set up neutral, reproducible benchmark experiments, compare forecasting methods - including foundation models and agents - under consistent evaluation protocols, and build their own benchmarking studies that can scale from local machines to larger compute environments. By the end of the session, attendees will have the practical knowledge and tooling needed to evaluate forecasting approaches on their own data, reproduce published results, and confidently select models based on evidence rather than hype. The workshop is intended for data scientists, machine learning practitioners, and researchers with basic Python knowledge who want a practical overview of the modern forecasting landscape and hands-on experience with the latest capabilities of the sktime ecosystem.