12-03, 21:00–21:30 (UTC), Data/ Data Science Track
Beneath the buzz of AI breakthroughs, a quiet revolution is unfolding in the world of forecasting: foundational time series models. These models promise to change the game for operational forecasting, but don’t expect magic. You won’t suddenly become a stock market oracle just by throwing data at them.
In this talk, we’ll peel back the layers of these new time series models, starting with how they work and how they evolved from transformers. We’ll tackle the big problems of limited data and overhyped algorithms, and explore the real-world challenges that make or break forecasts (hint: human input matters).
From scaling infrastructure to balancing accuracy and complexity, this session will provide practical insights to help you master forecasting in this new era of foundational models. The future of forecasting isn’t just about having the coolest model—it’s about knowing when and how to use it. Ready to unlock the secrets of foundational time series models and become the forecasting hero your data deserves? Join me, and let’s dive in:
Talk Outline
- Intro to Time Series Forecasting and Use Cases (2 min)
- What are Foundational Time Series Models? (10 min)
- A journey from local → global → foundational models
- Practical suggestions on using foundational time series models (10min)
- Feature Engineering and Human-in-the-loop: Models can’t use features that they don’t know about
- Deploying to Production: You still need engineering and infrastructure around your
model.predict() - When should you train your own model? Pro tip: Start from the LOTSA dataset & open source models
- Model Stability: Building robust, and resilient models
- What’s Next? (2min)
- Time series as images? Time series as novels?
Talk material: https://github.com/ahad-s/foundational-time-series-forecasting-models--pydata-global-2024
No previous knowledge expected
Ahad Shoaib is a Lead Data Scientist in the Infrastructure Data Science team at Salesforce, where he is responsible for taking time series forecasting and machine learning models all the way from “ideation” to “production”. Ahad holds a bachelor's degree in Computer Science & Math from the University of Waterloo.