A monster pot looks like a discontinuous jump when you only see the hand-by-hand chip count. Zoom into the same stretch action-by-action — preflop raise, cbet, turn barrel, river shove — and it's a continuous escalation, not a fluke. That's the paper's point about high-frequency price data, ported to the felt: apparent "jumps" are often just coarse sampling of ordinary continuous variance. Drag the slider below — same underlying session, same random draw, only the sampling resolution changes — and watch a real bipower-variation jump test flip its verdict live.
r₁, r₂, …, rₙ. Two estimators of "total variation" are computed:
RV ≫ BV, that's the signature of a jump.
The actual test statistic (Barndorff-Nielsen & Shephard, ratio form) is
TQ is the tripower quarticity (a third estimator, using products of three adjacent
|returns|^(4/3), needed to estimate the variance of the test statistic itself) and
θ = π²/4 + π − 5 ≈ 0.609 is a constant from the asymptotic theory. Under the null of "no jumps,
just continuous diffusion," z is asymptotically standard normal, so
z > 1.96 flags a jump at the 5% level. This demo computes RV, BV, TQ and z from scratch,
live, on whatever returns your slider selects — nothing here is a lookup table.