Data Tech 2026

This talk builds on prior work presented at FASTCon 2026.

Time Series Foundation Models (TSFMs) are gaining traction in forecasting, but their real-world behavior is more nuanced than the hype suggests. In our experiments, we observe that performance varies across data regimes, forecast horizons, and retraining setups—with only small differences among top models.

This talk focuses on the why behind these results.

We connect architectural choices in models like Chronos/Chronos2, TimeGPT, and TiReX to their observed behavior—covering data efficiency, generalization, regime sensitivity, and stability.

Rather than asking “which model is best,” we provide a practical framework for understanding when and why each model works, helping practitioners make more informed, context-driven decisions.