Regimes are a useful fiction. There is no official register of them and no reliable way to know in real time that one has ended, but returns clearly are not drawn from one unchanging distribution: volatility clusters, correlations rise in crashes, and trend persistence comes and goes.
Traders usually classify regimes crudely by volatility level, trend direction, and dispersion between sectors. Formal methods such as hidden Markov models produce cleaner labels but rarely a bigger edge, because the labels are still only known with a lag.
Example: 2017 realised volatility on the S&P 500 was around 7% annualised; 2020 spent weeks above 60%. A position size calibrated on the first would have been eight times too large in the second, which is the practical reason regimes matter.
Related: regime-filter