When Theory Bends — Learning Without i.i.d.

What statistical learning theory says (and doesn’t say) about forecasting with real-world data Series: Forecasting Reality — Machine Learning in a Non-Stationary World (Part 2)

In Part 1, we exposed the uncomfortable truth: most machine learning models assume i.i.d. data, but time series data is anything but.

So how do we reconcile this? If the theory behind machine learning doesn’t apply cleanly to time series, are we flying blind? Or have researchers found ways to stretch the theory—bend it, break it, rework it—for data that evolves over time?

The answer: yes, but it's complicated.