Bipower Variation: Jump-Robust Realized Volatility Explained

High-frequency realized variance adds the square of every intraday log return. As sampling gets finer under a suitable semimartingale model, RV converges to quadratic variation. When the price path has both a continuous diffusion and jumps, that limit includes both integrated variance and the sum of squared jumps. RV is therefore a useful measure of total ex-post variation, but it does not by itself say how much came from continuous movement versus discontinuities.

Bipower variation changes the building block. Instead of squaring one return, it multiplies the magnitudes of two neighboring returns. In the limiting model, an isolated jump contaminates one return interval, while its adjacent return is typically a small diffusion increment. The product involving that jump shrinks as the observation interval becomes finer; neighboring diffusion-return products continue to accumulate the continuous variation. Barndorff-Nielsen and Shephard introduced this use of power and bipower variation to separate integrated variance from rare-jump variation under stated models (Barndorff-Nielsen & Shephard, 2004).

The target is ex-post integrated variance, \(\int_0^T \sigma_t^2\,dt\), not a forecast of next period’s volatility. The realized-volatility calculation guide explains the squared-return measure; this article focuses on why adjacent absolute-return products have a different jump response.

Consider a log-price process over a fixed window:

\[ X_t=X_0+\int_0^t b_s\,ds+\int_0^t \sigma_s\,dW_s+J_t, \]

where the first term after \(X_0\) is a finite-variation drift, \(\sigma_t\) is the instantaneous diffusion volatility, and \(J_t\) is a jump component. For this process, quadratic variation separates into

\[ [X]_T=\int_0^T \sigma_t^2\,dt+\sum_{0<t\leq T}(\Delta X_t)^2, \]

with \(\Delta X_t=X_t-X_{t-}\). The first component is continuous integrated variance; the second is jump variation. A jump’s contribution is its size squared, regardless of whether the jump is positive or negative. The quadratic-variation guide explains why this path quantity differs from the variance of a collection of static prices.

With \(n\) intraday log returns \(r_i=X_{i\Delta}-X_{(i-1)\Delta}\), where \(\Delta=T/n\), realized variance is

\[ RV_n=\sum_{i=1}^{n}r_i^2. \]

As the mesh shrinks under the usual semimartingale conditions, \(RV_n\) estimates the full quadratic variation, including jumps. A volatility report that labels RV as “continuous volatility” without a no-jump assumption therefore mixes two targets.

Let \(Z\) be standard normal and \(\mu_1=E|Z|=\sqrt{2/\pi}\). A common realized bipower variation convention is

\[ BV_n=\mu_1^{-2}\frac{n}{n-1}\sum_{i=2}^{n}|r_i||r_{i-1}| =\frac{\pi}{2}\frac{n}{n-1}\sum_{i=2}^{n}|r_i||r_{i-1}|. \]

The factor \(\pi/2\) normalizes the product of absolute Gaussian increments to the continuous variance scale. The factor \(n/(n-1)\) is a small-sample endpoint correction used in a common convention; some definitions omit it because it tends to one as \(n\) grows. State which convention you use when comparing results.

The sum uses overlapping adjacent pairs: return 2 is paired with return 1, return 3 with return 2, and so on. It does not classify individual observations as “jump” or “no jump.” Its asymptotic result instead relies on the fact that, under finite-activity jumps, only finitely many small neighborhoods contain a jump as the grid gets finer, while ordinary neighboring diffusion returns generate the continuous integrated-variance limit. The limit theory allows broad stochastic-volatility behavior, including leverage-type dependence under its assumptions (Barndorff-Nielsen et al., 2006).

<!-- learn:illustration -->

Suppose one hypothetical session has seven log returns, in basis points:

Treat the fourth return as containing one hypothetical \(+12\) bp jump, with a negligible continuous move in that bin. Then

\[ RV_7=1+1+0.01+144+0.01+1+1=148.02\ \mathrm{bp}^2. \]

The six adjacent absolute-return products sum to \(4.60\ \mathrm{bp}^2\), so under the stated endpoint-corrected convention

\[ BV_7=\frac{\pi}{2}\frac{7}{6}(4.60)\approx 8.42994\ \mathrm{bp}^2, \qquad RV_7-BV_7\approx139.59006\ \mathrm{bp}^2. \]

The inserted jump’s squared size is \(12^2=144\ \mathrm{bp}^2\). The difference is not exactly 144: the adjacent products next to the jump include \(12\times0.1\) terms, and a seven-return sample also has ordinary finite-sample variation. This small invented example makes the arithmetic auditable; it is not observed market data or a claim that a short-window difference will accurately recover jump variation.

The standard consistency result is conditional. In a continuous stochastic-volatility semimartingale with a finite number of jumps over the window, observed at increasingly fine intervals without market microstructure noise, BV converges to the continuous integrated variance. In the same setup, \(RV_n-BV_n\) converges to the sum of squared jumps. This is an asymptotic separation, not a guarantee for every return series or every finite grid.

The finite-activity condition matters. With many small jumps or infinite jump activity, adjacent absolute-return products can receive non-negligible contributions from jump variation; the standard BV result need not isolate only the Brownian component. Extensions, truncation, or different jump tests may be needed, each with its own assumptions. Aït-Sahalia and Jacod develop a distinct high-frequency jump test with broader semimartingale scope; it is an alternative testing framework, not a proof that ordinary BV handles every jump process (Aït-Sahalia & Jacod, 2009).

Observation design also matters. Returns must refer to a common window and a defensible price series. Sparse intervals, stale quotes, asynchronous trading, overnight gaps, and jumps near session boundaries weaken a clean fine-grid interpretation. Dense sampling only helps if the observations approach the latent efficient price rather than adding more microstructure noise.

Under the finite-activity, noise-free asymptotic model, RV minus BV estimates jump variation. At a finite sample size, the difference can be negative even though squared jump variation cannot be negative. That can happen because RV and BV are noisy estimates of different limiting quantities. Do not silently clip negative values to zero: clipping changes the estimator’s sampling behavior and should be reported as an explicit rule.

A statistical test asks a different question: is the observed gap large relative to its sampling uncertainty under a no-jump null? Barndorff-Nielsen and Shephard derive asymptotic tests based on realized power variation and bipower variation; feasible versions estimate the relevant quarticity, often with multipower measures, rather than treating the raw difference as a significance threshold (Barndorff-Nielsen & Shephard, 2006). Threshold bipower variation was later developed to reduce finite-sample jump distortion and support inference under jump alternatives (Corsi, Pirino & Reno, 2010).

A detected jump is still a model-based statistical classification at a chosen sampling frequency and significance level. It does not identify a news cause, prove that a trader could have acted at the recorded price, or establish a profitable strategy.

Classical BV is not a noise-robust high-frequency estimator. With observed price \(Y_{t_i}=X_{t_i}+\epsilon_{t_i}\), bid–ask bounce, rounding, and quote updates add noise to each return. Since BV multiplies neighboring absolute returns, the noise does not disappear merely because the formula uses adjacent products. At very fine sampling, more observations can make the noise contribution more important.

This is a different problem from jumps. Realized kernels and two-scale realized variance target volatility estimation under microstructure noise with different constructions; they are not interchangeable with ordinary BV, and noise robustness alone does not imply jump robustness. The realized-kernel guide explains one of those noise-focused methods.

If the question requires handling both noise and jumps, use a method derived for both features and state its conditions. Podolskij and Vetter develop modulated bipower variation for noisy observations and construct versions robust to finite-activity jumps; that is a modified estimator, not the unadjusted formula above (Podolskij & Vetter, 2009). Practical alternatives may include less-frequent sampling, pre-averaging, thresholded measures, or a noise-robust estimator. Each changes the estimand, tuning choices, or uncertainty calculation.

BV estimates a historical continuous integrated-variance component under its assumptions. RV estimates total quadratic variation in a jump process. Their difference can estimate jump quadratic variation asymptotically. None of these quantities is automatically a forecast, a risk limit, or an entry signal. A forecasting model may use past RV and BV as separate predictors, but predictive value has to be evaluated out of sample with information timing and trading costs controlled.

The square root of a variance estimate is a volatility measure only after its time window and unit are made clear. If returns are in basis points, a variance is in bp² and its square root is bp over the sampled window. Annualization requires an explicit scaling assumption; multiplying by a calendar factor does not remove jumps, microstructure noise, or overnight measurement choices. The jump-risk guide discusses why discontinuous price changes also matter for hedge exposure, but it does not turn a historical BV decomposition into a trading instruction.

Name the asset, price input (trades, quotes, or midpoint), return definition, sampling frequency, session boundaries, and treatment of overnight intervals and missing observations. Report the RV and BV formulas, the finite-sample normalization convention, the number of returns, and the units. If you calculate \(RV-BV\), show the raw difference and state whether any threshold, truncation, or nonnegative clipping was applied.

Explain which model supports the intended interpretation: continuous stochastic volatility with finite-activity jumps, or a broader jump model with a different estimator. Disclose any microstructure-noise correction, jump-test statistic, quarticity estimate, significance level, and multiple-testing adjustment. Keep the uncertainty measure alongside the decomposition.

Bipower variation is useful because adjacent absolute returns retain a continuous-variation limit while reducing the asymptotic effect of isolated finite-activity jumps. Its value comes with strict boundaries: the standard formula is not a universal jump filter, a significance test, or a microstructure-noise correction. Keep those boundaries attached to any risk or research use.