Agentic Forecasting with sktime: Towards Autonomous Time Series Analysis

Recent advances in LLM and AI agents are transforming how practitioners interact with forecasting systems. Rather than manually selecting models, tuning hyperparameters, and interpreting results, users can increasingly delegate parts of the forecasting workflow to intelligent agents that reason about data, choose appropriate methods, execute experiments, and explain their findings.

This talk introduces the emerging paradigm of agentic forecasting in the sktime ecosystem. We present how language-model-driven agents can automate key forecasting tasks - from understanding forecasting problems and constructing analysis workflows to selecting models, running experiments, and interpreting results - while remaining grounded in reproducible and transparent machine learning practices.

We showcase how agentic capabilities integrate with sktime’s unified forecasting framework and model ecosystem, enabling users to accelerate experimentation without sacrificing rigor. We also discuss the opportunities and challenges of combining foundation models, domain knowledge, and reproducible evaluation to create trustworthy AI assistants for time series forecasting.

The talk concludes with a vision for the future of forecasting, where human expertise and AI agents work together to make advanced time series analysis more accessible, efficient, and reliable.