Seonglae ChoA property where reordering data does not change its probabilistic characteristics. → The assumption that "when" something was observed doesn't matter, only "what" was observed. i.i.d. ⇒ always exchangeable, but exchangeable ⇒ not necessarily i.i.d. A weaker assumption than i.i.d. Exchangeability can hold even with dependencies (e.g., latent variable structures).
Conformal Prediction
Relies on data symmetry (exchangeability) rather than models. Calculate a nonconformity score for each sample (how unusual it is). The score of new data has an equal position among existing scores → rank follows a uniform distribution → use this to generate distribution-free prediction intervals. Works with any model, without distributional assumptions → guarantees finite-sample coverage
