Reader question: A day can be volatile because prices moved continuously for hours, or because one sudden jump contributed a large fraction of total variation. Can high-frequency returns separate those two mechanisms without first fitting a parametric jump-diffusion model?

Realized bipower variation (BPV) was developed for exactly that measurement problem. It uses products of adjacent absolute high-frequency returns to estimate the continuous integrated variance while being robust, under its classical assumptions, to infrequent jumps. Comparing ordinary realized variance with BPV then gives a model-light way to isolate the discontinuous contribution to quadratic variation.

This article owns one precise computational job: jump-robust ex-post decomposition of high-frequency quadratic variation into continuous and discontinuous components. It does not own noise-robust realized variance, OHLC volatility, stochastic-volatility forecasting or parametric jump-diffusion pricing.

This is public mathematical and computational education. It is not a trading signal, financial advice or a claim that every large intraday move is an economically meaningful jump.

1. Realized variance measures total quadratic variation

Let log prices be observed at intraday times and define returns:

r_i = p_{t_i} − p_{t_{i−1}}.

The realized variance is:

RV = Σ r_i².

For a semimartingale price process with continuous volatility plus jumps, sufficiently fine noise-free sampling makes RV converge to total quadratic variation:

[p,p]_T = ∫₀^T σ_t²dt + Σ_0<t≤T(Δp_t)².

The first term is continuous integrated variance. The second is jump variation. RV combines them.

2. The bipower construction

For standard normal Z, define:

μ₁ = E|Z| = √(2/π).

A common realized bipower-variation estimator is:

BPV = μ₁⁻² Σ_i=2ⁿ |r_i||r_i−1|.

Because μ₁⁻² = π/2, this is often written:

BPV = (π/2) Σ |r_i||r_i−1|.

Normalisation conventions can include small finite-sample adjustments. A production implementation must record the exact convention used.

3. Why one jump affects RV more than BPV

Suppose one return contains a jump J and dominates its interval.

RV receives approximately:

J²

from that observation.

BPV instead multiplies the jump-contaminated absolute return by neighbouring returns. Under high-frequency sampling, those neighbours shrink with the sampling interval. The jump therefore does not survive in BPV in the same way it survives in squared-return variation.

That is the core robustness mechanism: squaring a jump preserves it; multiplying it by an adjacent diffusion-scale return makes its contribution asymptotically negligible under the classical finite-activity setting.

4. Continuous variation estimate

Under suitable regularity conditions, BPV converges to:

IV = ∫ σ_t²dt.

This makes BPV a nonparametric estimator of the continuous component of daily price variation.

Then the natural jump-variation estimate is:

JV ≈ RV − BPV.

In finite samples, RV−BPV can be slightly negative because both estimators are noisy. A reporting layer may show the raw difference and, separately, a nonnegative truncated diagnostic such as max(RV−BPV,0). The truncation rule should never be hidden.

5. A simple numerical illustration

Imagine most one-minute returns are around ±0.10%, but one news interval contains a +2.00% move.

The jump interval contributes:

0.02² = 0.0004

to RV.

In BPV, the same 2% absolute return is multiplied by neighbouring diffusion-scale returns such as 0.001. That produces terms around:

0.02 × 0.001 = 0.00002

before normalisation—far smaller than the squared jump contribution.

The example is only intuition; the formal robustness statement is asymptotic.

6. The BNS jump-testing idea

Barndorff-Nielsen and Shephard developed formal tests of whether observed quadratic variation is consistent with a continuous sample path.

The basic information is:

RV − BPV.

But a raw difference is not enough. It must be scaled by an estimate of its sampling variance, which depends on integrated quarticity. Multipower measures such as tripower quarticity are commonly used for that normalisation because they are themselves designed to remain robust to jumps.

The resulting standardized statistic can be compared with an asymptotic reference distribution.

7. Why quarticity enters

Variance of a variance estimator depends on fourth-order volatility quantities.

If the local diffusion variance is σ_t², the relevant integrated fourth moment has the form:

∫ σ_t⁴dt.

A jump test that ignores this scale can interpret ordinary sampling noise as a jump on high-volatility days and miss meaningful discontinuities on quieter days.

8. Inputs and outputs

Inputs can include:

clean intraday log prices or returns;
sampling interval;
session boundaries;
corporate-action adjustments;
bad-tick and outlier rules;
overnight-return policy;
BPV normalisation convention;
quarticity estimator;
jump-test significance level.

Outputs can include:

realized variance RV;
bipower variation BPV;
raw jump-variation difference RV−BPV;
continuous-variation share;
jump-variation share;
standardized BNS-style test statistic;
jump flag;
sampling-frequency sensitivity;
data-quality flags;
comparison with alternative jump tests.

9. Evidence polarity

Evidence for confidence includes:

RV and BPV are close on visually continuous days;
large RV−BPV gaps coincide with independently documented price discontinuities;
jump flags are stable across nearby reasonable sampling intervals;
simulation under known diffusion-plus-jump processes recovers continuous and jump components;
quarticity-normalized statistics have reasonable finite-sample size;
results survive trade-versus-midquote and cleaning sensitivity checks.

Evidence against confidence includes:

jump flags explode as sampling becomes ultra-fine;
results reverse after correcting one bad tick;
BPV rises sharply with bid–ask bounce;
daily tests label implausibly large fractions of days as jump days;
small changes in sampling frequency change the sign of RV−BPV repeatedly;
overnight gaps are mixed inconsistently with intraday variation.

10. Microstructure noise is a major weak link

The classical BPV theory is not a license to use every raw transaction tick.

Observed prices contain bid–ask bounce, discreteness, timestamp issues and other market-microstructure noise. At very high frequency, noise can dominate efficient-price changes.

Both RV and BPV can then be distorted.

Falsifier: compute the decomposition across a frequency grid. If the continuous and jump estimates change materially as one moves from ultra-high-frequency ticks to modestly sparse sampling, the noise problem is unresolved.

See Realized-Kernel Algorithms for the separate noise-robust variance problem.

11. A single bad tick can look exactly like a jump

Suppose a price feed records 100, then 120, then 100 because of one erroneous observation.

RV records two enormous squared returns. BPV records large adjacent products as well, but the decomposition can still indicate unusual discontinuous variation.

No statistical jump test can infer from prices alone whether the event was genuine news, a bad print or a data merge error.

Falsifier: reconcile every extreme detected event against raw trade/quote records and data-quality flags before assigning economic meaning.

12. Finite-activity versus infinite-activity jumps

The clean intuition is strongest for infrequent, finite-activity jumps.

Processes with many small jumps can blur the distinction between “continuous” and “jump” variation at practical sampling frequencies. Different power-variation methods have different asymptotic behaviour under finite- and infinite-activity jump processes.

Boundary: a BPV decomposition is not a universal classifier for every Lévy or rough jump structure.

See Variance-Gamma Algorithms for a pure-jump model where jump activity is structurally different from a rare-jump diffusion.

13. Daily multiple-testing can create too many “jump days”

If one test is performed every day for years, even a correctly sized test generates false rejections.

Finite-sample studies also show that realized jump tests can have nontrivial size and power distortions.

Falsifier: report the expected false-rejection count implied by the significance level, examine long-span behaviour and compare multiple test families rather than treating one daily rejection as ground truth.

14. The event window matters

A price adjustment that occurs over several consecutive intervals may be economically sudden but statistically spread out.

One large return produces a clear RV contribution. Ten moderately large adjacent returns can also produce substantial BPV because adjacent products are now large.

So BPV is not a semantic “news detector.” It distinguishes mathematical path behaviour under specific high-frequency asymptotics.

15. Overnight returns are a separate design choice

Many BPV applications use intraday returns and exclude the close-to-open return because it spans a long period with no continuous observation.

If the research question is full close-to-close variance, overnight variation must be incorporated under a documented convention.

Falsifier: compare intraday-only and close-to-close decompositions. A large difference means the overnight information channel is economically material and should not be silently dropped.

16. Jump magnitude versus jump significance

A positive RV−BPV difference is an estimated magnitude. A BNS-style standardized statistic is a significance test.

Those are different objects.

A small jump contribution can be statistically significant on a quiet day; a larger raw difference can be statistically ambiguous on an extremely volatile day.

A useful output should show both estimated jump variation and statistical evidence.

17. Alternatives

Aït-Sahalia–Jacod power-variation tests: compare high-order power variation across sampling scales and have different robustness properties.

Threshold/truncated realized variation: classify unusually large returns relative to a local volatility threshold and exclude them from the continuous estimator.

Pre-averaged or noise-robust jump tests: designed when microstructure noise is explicitly part of the observation model.

Parametric jump models: estimate jump intensity and size jointly with diffusion parameters, at the cost of stronger model assumptions.

18. Relation to realized kernels and Yang–Zhang

Realized kernels ask how to estimate variation when high-frequency observations are noisy.

Yang–Zhang volatility uses daily OHLC information and treats overnight movement explicitly.

BPV owns neither problem. Its canonical question is continuous versus jump variation.

19. Relation to Hayashi–Yoshida

Hayashi–Yoshida covariance solves nonsynchronous cross-asset covariance estimation. Bipower variation is primarily a univariate jump-robust variation measure, though multipower and multivariate extensions exist.

21. What would falsify confidence?

Confidence should be withdrawn if simulation under a continuous model produces excessive rejections; if a known jump simulation is rarely detected; if jump estimates are dominated by data errors; if frequency changes reverse conclusions; if noise-robust alternatives disagree systematically; or if the decomposition cannot be reproduced from the archived return series and normalization rules.

22. Verification and update triggers

Preserve the raw-data identifier, price type, sampling grid, cleaning policy, BPV convention, quarticity estimator, significance threshold, overnight policy and all sensitivity outputs.

Revalidate when:

market-data feed changes;
tick size changes;
liquidity regime changes;
sampling frequency changes;
overnight/session definitions change;
the target changes from rare jumps to infinite-activity dynamics;
the jump-test family changes.

23. Primary and high-quality references

Ole E. Barndorff-Nielsen and Neil Shephard, Power and Bipower Variation with Stochastic Volatility and Jumps, Journal of Financial Econometrics, 2004.
Barndorff-Nielsen and Shephard, Econometrics of Testing for Jumps in Financial Economics Using Bipower Variation, Journal of Financial Econometrics, 2006.
Oxford University Research Archive, institutional record for Power and Bipower Variation.
Daisuke Kurisu, Power Variations and Testing for Co-Jumps: The Small Noise Approach, Scandinavian Journal of Statistics, 2018.

Educational boundary: Bipower variation is a statistical decomposition of observed path variation under explicit asymptotic assumptions. A detected discontinuity is evidence about price-path structure, not proof of a specific cause or a prediction of future returns.

Small Group Tutorials

How Realized-Bipower-Variation Algorithms Separate Continuous Volatility from Jumps: Adjacent Absolute Returns, Integrated Variance, BNS Tests, Microstructure Noise and Jump-Detection Failure