changepoint-detector had three volatility options tuned only on simulated noise. This run tried them on 1,750 real days of exchange-rate returns (public-domain Fed data), where the 2015 Swiss franc floor removal and the 2016 Brexit vote are known dates.
All four modes find both events. They differ on everything else: on the franc the default reports six changes, --robust and --dispersion three; on the pound --dispersion leaves three of five. --bootstrap, which cut false alarms from 51% to 16% on simulated GARCH noise, returned the same six franc dates as the default. That is one series, so it is a warning rather than a verdict.
Reproduce with node eval/fx-shocks.mjs in the venture. Statistical analysis of public data, not financial advice.