High-Frequency Financial Econometrics

Barndorff-Nielsen & Shephard (2004, 2006) Bipower Variation & Jump Test Lab

Deconstruct intraday market volatility into continuous diffusive variance and discontinuous jump risk using Ole Barndorff-Nielsen and Neil Shephard's Realized Bipower Variation ($BV$) and the asymptotic $Z$-test for discrete jumps.

Seminal Papers: J. Fin. Econometrics (2004) & Econometrica (2006) • Core Estimator: Bipower Variation ($BV$) • Hypothesis Test: Asymptotic Normal $Z$-statistic

Continuous Diffusion vs. Jump Discontinuities

High-frequency asset logarithmic prices $p_t = \ln S_t$ follow a continuous jump-diffusion semimartingale:

$$dp_t = \mu_t dt + \sigma_t dW_t + J_t dN_t$$ $$RV_t = \sum_{j=1}^M r_{t, j}^2 \xrightarrow{p} \int_{t-1}^t \sigma_s^2 ds + \sum_{s \in (t-1, t]} J_s^2$$

Total Realized Variance ($RV$) converges to the sum of integrated continuous variance plus the quadratic variation of all discrete jumps $J_s$ occurring during the day.

To isolate continuous integrated variance, Barndorff-Nielsen & Shephard (2004) proved that multiplying adjacent absolute returns eliminates isolated jumps:

$$BV_t = \frac{\pi}{2} \left(\frac{M}{M-1}\right) \sum_{j=2}^M |r_{t, j}| |r_{t, j-1}| \xrightarrow{p} \int_{t-1}^t \sigma_s^2 ds$$ $$JV_t = \max(RV_t - BV_t, 0), \quad RJ_t = \frac{JV_t}{RV_t}$$

Because a discrete jump appears in only one interval $r_j$, its product with adjacent continuous increments is of order $O_p(M^{-1/2})$, vanishing asymptotically as $M \to \infty$.

Sampling & Jump Controls

78 = 5-min intervals, 390 = 1-min intervals
Total Realized Vol ($RV$)

62.8%

$\sqrt{252 \times RV}$
Continuous Vol ($BV$)

21.9%

$\sqrt{252 \times BV}$
Disentangled Jump Vol

58.8%

$\sqrt{252 \times JV}$
Relative Jump Ratio ($RJ$)

87.8%

$JV / RV$
BNS Test Statistic ($Z$)

9.84

Asymptotic $\mathcal{N}(0, 1)$
P-Value ($p$)

< 0.001

$1 - \Phi(Z)$
Statistical Decision

Significant Jumps (p < 0.01)

Rejection of Pure Diffusion
Tri-power Quarticity ($TQ$)

4.82

$\int_0^1 \sigma_s^4 ds \ (\times 10^{-6})$
High-Frequency Econometric Interpretation

When $Z > 1.96$ ($p < 0.05$), the null hypothesis of pure continuous Gaussian diffusion is formally rejected. Realized Bipower Variation successfully filters out discrete shocks, permitting accurate estimation of diffusive integrated volatility without contamination from jump variance.

Figure 1: Intraday high-frequency simulated price trajectory. Discrete jump events are flagged with prominent markers (green for positive jumps, red for negative shocks). Notice how the price exhibits continuous Brownian motion interspersed with sharp discrete discontinuities.

Figure 2: Interval-by-interval comparison of squared returns $r_j^2$ (red) versus Bipower Variation products $\frac{\pi}{2}|r_j||r_{j-1}|$ (cyan). When a jump occurs, $r_j^2$ spikes massively, whereas the adjacent Bipower product remains small because the jump is dampened by its neighbor's continuous scale.

Figure 3: Volatility Signature Plot evaluating $RV$ (red) and $BV$ (cyan) across sampling intervals from 1 minute to 65 minutes. At ultra-high frequencies (1-min), microstructure noise may inflate variance, whereas at intermediate frequencies (5 to 15-min), $BV$ stabilizes cleanly around the true continuous diffusive benchmark.

Sensitivity Grid 1: BNS Test Statistic $Z$

Asymptotic jump test statistic $Z$ across continuous volatility $\sigma_{\text{cont}}$ and discrete jump magnitude $\Delta J$. ($Z > 1.96$ indicates statistical significance).

Sensitivity Grid 2: Relative Jump Contribution ($RJ$ %)

Fraction of total day variance driven by discrete jumps $RJ = \frac{JV}{RV} \times 100\%$ across intraday sampling intervals $M$ and jump size $\Delta J$.

Asymptotic Theory of Bipower Variation and Jump Tests

1. The Asymptotic Robustness of Bipower Variation

In standard diffusion models, squaring return increments $r_j^2 = (p_{j\Delta} - p_{(j-1)\Delta})^2$ guarantees that as $\Delta \to 0$, the sum converges to quadratic variation $[p, p]_t$. However, if prices exhibit jumps:

$$[p, p]_t = \int_0^t \sigma_s^2 ds + \sum_{0 < s \le t} (\Delta p_s)^2$$

Ole Barndorff-Nielsen and Neil Shephard recognized that for an isolated jump at time $\tau \in ((j-1)\Delta, j\Delta]$, the adjacent interval $|r_{j-1}|$ contains only continuous Brownian motion of order $O_p(\Delta^{1/2})$. Consequently, the cross-product $|r_j||r_{j-1}| \approx |J_\tau| \cdot O_p(\Delta^{1/2}) \to 0$ as $\Delta \to 0$, filtering out jumps completely.

2. The BNS Linear & Log-Ratio Jump Tests

Barndorff-Nielsen & Shephard (2006) and Huang & Tauchen (2005) derived the joint asymptotic distribution of $(RV, BV)$. Under the null hypothesis $H_0: \text{No Jumps}$, the relative difference converges to Gaussian noise:

$$Z = \frac{\frac{RV - BV}{RV}}{\sqrt{\left[\left(\frac{\pi}{2}\right)^2 + \pi - 5\right] \frac{1}{M} \max\left(1, \frac{TQ}{BV^2}\right)}} \xrightarrow{d} \mathcal{N}(0, 1)$$

Tri-power Quarticity ($TQ$) scales the asymptotic variance by integrated quarticity $\int_0^1 \sigma_s^4 ds$. The max adjustment $\max(1, TQ/BV^2)$ provides robust finite-sample size control against microstructure noise.

Self-Assessment Mastery Quiz

Verify your mastery of Realized Bipower Variation, jump-diffusion disentanglement, and the BNS non-parametric hypothesis test.

1. What quantity does Realized Bipower Variation ($BV$) isolate in high-frequency price data?

2. Why is the multiplying factor $\frac{\pi}{2} \approx 1.5708$ applied in the Bipower Variation definition?

3. If a trading day has $RV = 0.0004$ and $BV = 0.0001$, what is the estimated Disentangled Jump Variation ($JV$)?

4. What does a BNS test statistic of $Z = 3.45$ ($p < 0.001$) imply?

Answer all 4 questions above to evaluate your mastery score.

Related Volatility & Microstructure Simulators

Yang-Zhang Range Volatility Lab

Estimate historical volatility with 8x to 14x greater efficiency using Dennis Yang and Qiang Zhang's (2000) jump- and drift-adjusted OHLC model.

MRR Spread Decomposition Lab

Deconstruct price changes into adverse selection trade surprises, order processing friction, and public news shocks.

Order Flow Imbalance (OFI) Lab

Decompose limit order book event transitions and quantify contemporaneous high-frequency price impact.

GARCH(1,1) Volatility Forecasting Lab

Model volatility clustering, mean-reversion persistence, unconditional variance, and term structure forecasts.

Merton Jump-Diffusion Lab

Value options under compound Poisson jumps, lognormal jump sizes, and fat-tailed volatility smiles.

Frequently Asked Questions

Realized Bipower Variation (BV) is a high-frequency econometric estimator introduced by Ole Barndorff-Nielsen and Neil Shephard (2004). By summing the product of consecutive absolute intraday returns multiplied by pi/2, BV isolates purely continuous integrated variance while remaining asymptotically immune to discrete price jumps.

A discrete jump impacts only a single return interval r_j. Because BV computes the product of adjacent absolute returns |r_j| * |r_{j-1}|, the single jump is multiplied by an infinitesimal continuous return of order O_p(M^{-1/2}), vanishing asymptotically as the sampling frequency M approaches infinity.

The BNS test evaluates the standardized difference between total Realized Variance (RV) and continuous Bipower Variation (BV). Under the null hypothesis of no jumps, the test statistic Z converges to a standard normal distribution N(0, 1), allowing formal statistical rejection of pure diffusion when Z exceeds critical thresholds.

Realized Tri-power Quarticity (TQ) is a jump-robust estimator of integrated quarticity (the integral of sigma^4). It scales the product of three consecutive absolute returns raised to the 4/3 power to standardize the asymptotic variance denominator of the BNS test statistic.

Continuous volatility represents predictable, hedgeable diffusive risk that can be dynamically neutralized with Black-Scholes delta hedging. In contrast, discrete jump variation represents unhedgeable tail risk, gap risk, and liquidity collapse that requires non-linear options insurance or regulatory capital reserves.