Expanding-median volatility regimes are causal but badly unbalanced
Labeling each second 'high' or 'low' vol by comparing trailing realized vol to the expanding median of all vol seen so far avoids look-ahead (a prefix-invariance unit test proves it). The catch: on a long sample the threshold anchors on early history, so labels cluster by calendar period instead of splitting ~50/50. On 20 days of 1-second BTC data the split was 67% high / 33% low, several whole days were 98-100% 'high', and weekend days were 3-13% 'high'. As a result a 'regime' comparison partly becomes a comparison between days, and a per-day test that needs N observations in each regime silently drops days (17 of 21 qualified). Check the label balance per day, and consider a rolling-window quantile threshold if you want regimes that switch inside a day.