Cryptocurrency markets exhibit rapid shifts in volatility that require specialized statistical tools to detect regime changes. Empirical finance uses several complementary tests and models that identify abrupt breaks, persistent regime shifts, or latent state changes in volatility, informing risk management and policy responses.
Statistical methods for abrupt breaks and volatility clustering
The ARCH LM test developed by Robert F. Engle New York University detects autoregressive conditional heteroskedasticity and is a first-step indicator of volatility clustering; significant ARCH effects suggest time-varying volatility that may hide regime changes. For explicit structural breaks, the Bai–Perron structural break test by Jushan Bai Columbia University identifies multiple breakpoints in linear models and estimates their dates, useful when regulatory events or exchange outages produce discrete shifts. CUSUM and CUSUMSQ tests detect parameter instability over time and can flag periods where volatility dynamics change materially. Classical variance-comparison tests such as Levene or Bartlett tests can detect shifts in unconditional variance across pre-specified windows, but they are sensitive to non-normal returns and heavy tails typical of crypto data.
Regime-switching and change-point approaches
When regimes are latent rather than tied to known dates, Markov-switching models introduced by James D. Hamilton University of California San Diego capture switches between low- and high-volatility states; estimation via maximum likelihood or expectation-maximization permits inference on transition probabilities and state durations. Extensions like MS-GARCH combine regime-switching with conditional heteroskedasticity to model persistence within regimes. Nonparametric change-point detection methods, including Bayesian change-point models and algorithms such as Pruned Exact Linear Time, provide flexible detection without strong parametric assumptions and can be adapted for online monitoring.
Model selection and inference often rely on likelihood-ratio, sup-Wald or sup-LR statistics and information criteria such as AIC and BIC, with bootstrap methods commonly used to obtain robust critical values under non-standard distributions. Practical challenges include small-sample bias, exchange-level microstructure noise, and cross-border liquidity differences; these human and territorial nuances mean that identical tests can yield different break dates across exchanges or regulatory jurisdictions. Detecting regime changes in cryptocurrency volatility therefore combines tests for conditional heteroskedasticity, structural breaks, and latent regimes, with corroboration across methods recommended to guard against spurious findings and to guide trading and regulatory decisions.