l5-fooled-by-randomness_stationarity-problem

The stationarity problem is the issue that statistical methods assume the process generating data is stable, but in real-world systems the process itself changes. Drawing conclusions from past data fails when the underlying generator has been altered.

Extrapolation fails because the market of one era is not the same generator as another era. The rules, participants, and structure change, so probabilities computed from old data describe a game that no longer exists.

The urn metaphor contrasts textbook probability, where an urn has fixed known composition, with reality, where children add, remove, and repaint balls while you draw. You unknowingly compute probabilities for a changing system.

Distrust them when the system under study is social, economic, or political, where the generating process is in flux. One realization from one past generator gives no guarantee the generator has not been swapped.