Microstructure Noise-Robust Jump Test

• The paper introduces tests that maintain statistical validity and high sensitivity by adjusting for different models of microstructure noise in high-frequency financial prices.
• It employs methods such as blockwise local minima, pre-averaging, and plug-in de-noising to isolate jump components from noise effects.
• Empirical findings show that noise-adapted jump tests outperform traditional methods, offering improved detection rates and robust volatility estimation.

A microstructure noise-robust jump test is a statistical procedure to detect and infer discontinuities (jumps) in high-frequency financial price data when such data are contaminated by market microstructure noise. Microstructure noise—arising from frictions such as discreteness of prices, bid/ask bounce, and order book dynamics—impedes direct jump detection, especially at the ultra-high frequencies prevalent in limit order book data. The central challenge is to develop tests that maintain statistical validity and high sensitivity in the presence of non-classical, possibly one-sided or parametric, noise.

1. Statistical Models for Noisy High-Frequency Prices

High-frequency order price models represent the observed (log-)prices $Y_i$ as a combination of an underlying efficient price process $X_t$ and a noise component. The efficient price is typically modeled as an Itô semimartingale allowing for both continuous and jump components: $$X_t = X_0 + \int_0^t a_s\,ds + \int_0^t \sigma_s\,dW_s + \int_0^t\int_{|z|\le1} \delta(s,z)\,(\mu-\nu)(ds,dz) + \int_0^t\int_{|z|>1} \delta(s,z)\,\mu(ds,dz)$$ Here, $a_t$ is locally bounded drift, $\sigma_t > 0$ the spot volatility, $W_t$ a Brownian motion, $\mu$ a Poisson random measure, and $\delta(s,z)$ the jump kernel.

Several observation noise models are pertinent for microstructure contamination:

• One-sided microstructure noise (LOMN): $Y_i = X_{t_i^n} + \epsilon_i$, $\epsilon_i \ge 0$, with cumulative distribution $F_{\eta}(x) = \eta x (1 + o(1))$ as $x \downarrow 0$. No further moment or tail constraints are imposed (Bibinger et al., 2024).
• Centered additive noise (MMN): $Y_i = X_{i/n} + \varepsilon_i$, $\varepsilon_i$ centered, possibly serially correlated and endogenous, with bounded moments and mixing conditions (Bibinger et al., 2014).
• Parametric microstructure noise: $Z_{t_i}=X_{t_i}+\phi(Q_{t_i},\theta_0)$, where $Q_{t_i}$ are limit order book covariates, and $\phi(\cdot,\theta_0)$ is a known impact function parametrized by $\theta_0$ (Clinet et al., 2017).

The specifics of the noise process—symmetry, support, and structure—profoundly impact the construction and optimality of noise-robust jump tests.

2. Methodologies for Noise-Robust Jump Detection

Noise-robust jump tests are designed to preserve statistical power and control type I error in the presence of microstructure noise. The main methodological approaches include:

2.1 Local Order Statistic–Based Jump Test (LOMN Model)

For one-sided noise, (Bibinger et al., 2024) introduced a global jump test using blockwise local minima:

• Partitioning: Divide the time interval $[0,1]$ into blocks of length $h_n$; $$\mathcal{I}_k^n = \{i: t_i^n \in (kh_n, (k+1)h_n)\}$$.
• Blockwise minima: $m_{k,n} = \min_{i \in \mathcal{I}_k^n} Y_i$.
• Volatility estimation: Use increments of block minima:

$$\hat\sigma^2_{\tau-} = \frac{\pi}{2(\pi-2)K_n} \sum_{r=(\lfloor h_n^{-1}\tau\rfloor-K_n)\vee1}^{\lfloor h_n^{-1}\tau\rfloor-1} h_n^{-1}(m_{r,n} - m_{r-1,n})^2$$

• Test statistic:

$$T^{BHR} = \max_{k=1,\dots,h_n^{-1}-1} \left| \frac{m_{k,n}-m_{k-1,n}}{\hat\sigma_{k h_n}} \right|$$

where $\hat\sigma_{k h_n}$ may use a truncated estimator to suppress jump-induced outliers.

2.2 Pre-averaging and Smoothing–Based Tests

Tests for additive noise employ smoothing techniques to attenuate the impact of noise:

• Lee–Mykland (Pre-Averaging) Test (“LM12”): Divides the sample into blocks, computes blockwise averages $\bar P_j$, then forms block returns $Y_j$. The test statistic is the standardized maximum $|Y_j|$; critical values are derived from Gumbel extreme-value theory (Maneesoonthorn et al., 2017).
• Aït-Sahalia–Jacod–Li (Kernel Smoothing) Test (“ASJL”): Constructs a kernel-smoothed path and applies power variation statistics to the smoothed process. The central limit theorem is used for critical values (Maneesoonthorn et al., 2017).

2.3 Plug-in De-noising with Parametric Noise

When the microstructure noise arises from known parametric mechanisms:

• Estimate noise parameters (e.g., via MSE minimization):

$$\widehat\theta = \arg\min_{\theta} \sum_{i=1}^N (\Delta_i Z - [\phi(Q_{t_i},\theta) - \phi(Q_{t_{i-1}},\theta)])^2$$

• Construct de-noised prices: $$\widehat X_{t_i} = Z_{t_i} - \phi(Q_{t_i},\widehat\theta)$$.
• Apply standard jump tests (e.g., bipower-variation, truncation-based) on $\widehat X_{t_i}$, with all asymptotics valid under mild conditions (Clinet et al., 2017).

3. Asymptotic Theory, Null Distributions, and Thresholds

The limiting distribution of the noise-robust jump test under the global no-jump hypothesis and the derivation of rejection thresholds are central features of modern methodology.

• LOMN Model (Blockwise Minima): Under $H_0: \sup_t |\Delta X_t|=0$, the test statistic has an extreme-value limit:

$$n^{1/3}T^{BHR} - 2\log(2h_n^{-1}-2) + \log\big(\pi \log(2h_n^{-1}-2)\big) \xrightarrow{d} \Lambda$$

where $\Lambda$ is the standard Gumbel law.

• LM12: The test statistic converges to a Gumbel law $P(T \le x) = \exp\{-2e^{-x}\}$.
• ASJL: Under the null, the standardized statistic converges to $N(0,1)$.
• Plug-in Framework: The plug-in versions of classical jump tests preserve their original asymptotic null distributions (e.g., normal or chi-square), provided the noise-parameter estimator converges sufficiently quickly (Clinet et al., 2017).

Critical values are computed from these limiting distributions, and in practice, nonparametric bootstrap methods may be used for refined finite-sample inference (Bibinger et al., 2024).

4. Detection Rates, Power, and Comparison Across Models

Detection ability—the minimal size of a jump that can be reliably detected—depends heavily on the noise structure.

• LOMN (One-sided): Capable of detecting jumps of size $O(n^{-\beta})$, $\beta < 1/3$, i.e., shrinking at any rate slower than $n^{-1/3}$. This is strictly better than the $n^{-1/4}$ rate attainable under classical MMN additive noise. The use of local minima exploits the support asymmetry (one-sided noise), yielding faster detection and increased sensitivity to small jumps (Bibinger et al., 2024).
• Additive Centered Noise: The optimal detection rate is $n^{-1/4}$ due to the variability introduced by symmetric additive noise (Bibinger et al., 2014).
• Plug-in Parametric Approach: Retains the consistency and power of standard noise-free jump tests, with only $O(N^{-1})$ bias contributed by the parametric estimation step.

Simulations confirm that methods adapted to one-sided or parametric noise outperform classical tests when these noise conditions hold, both in terms of size control and power. Smoothing-based pre-averaging and kernel tests (e.g., LM12, ASJL) perform optimally only if ultra-high-frequency data are available; their power diminishes as sampling frequency decreases (Maneesoonthorn et al., 2017).

5. Implementation Guidelines and Practical Considerations

The performance of noise-robust jump tests critically depends on tuning parameters, sampling frequency, and the nature of market microstructure contamination.

• Block Length and Truncation: For blockwise minima tests, $h_n = c n^{-2/3}$ (with $c \approx 1.2$) and $K_n \gg 1$ (e.g., $K_n = 200$) are recommended. Truncation is utilized to exclude large increments likely caused by jumps from volatility estimation.
• Bootstrap Procedures: Simulate under the estimated volatility and noise configuration to improve finite-sample test size and power.
• Empirical Strategy: Apply the test to each side of the limit order book separately (best ask/bid), compare with mid-quote variants under classical MMN to exploit differences in sensitivity.

A table condensing key implementation details from (Bibinger et al., 2024):

Step	Description	Typical Value/Method
Block length $h_n$	Partition for minima, set as $c n^{-2/3}$	$c \approx 1.2$
$K_n$	Number of past blocks for volatility estimation	$K_n = 200$
Truncation threshold	Cutoff for jump-robust volatility estimator	$u_n = \beta^{tr} h_n^\kappa$
Bias correction	Empirical factor for spot volatility	Multiplicative $\approx 0.954$
Noise-level estimate	For bootstrap, based on $\sqrt{2}\hat q_n$	See description

These choices ensure type I error control, robustness to microstructure configuration, and high jump-detection sensitivity.

6. Extensions, Empirical Findings, and Related Test Families

Noise-robust jump tests have been extended to handle not only univariate jumps but also simultaneous price and volatility jumps, infinite activity jump processes, and nonparametric spot volatility estimation. Spectral methods and adaptive weighting improve finite-sample efficiency and reduce estimation error for local volatility under noise (Bibinger et al., 2014). Plug-in de-noising enables application of the full suite of semimartingale-based jump tests to noisy prices, with negligible asymptotic impact (Clinet et al., 2017).

Empirical studies on U.S. equity and NASDAQ order book data confirm:

• Substantial power gains and enhanced detection of small jumps using LOMN tests.
• Intra-daily analysis is feasible in real time without down-sampling or ad hoc filtering.
• For days with price jumps, a high proportion of volatility jumps are detected, with relative volatility-jump sizes averaging $23\%$–$37\%$.

A plausible implication is that tailoring the jump detection methodology to the structural properties of microstructure noise can yield significant improvements in empirical asset price modeling, risk management, and market microstructure analysis.

7. Limitations and Trade-offs

While LOMN-based and plug-in noise-robust jump tests deliver superior theoretical rates and empirical performance under their respective model assumptions, their optimality is not universal:

• For extremely sparse or irregularly spaced data, or noise departing substantially from model assumptions (e.g., heavy-tailed, dependent, or endogenous structures), performance may degrade.
• Pre-averaging and kernel smoothing tests require genuinely ultra-high frequency data for effectiveness.
• Plug-in methods depend on the correctness and completeness of the parametric noise model, and may inherit bias if limit order book covariates misspecify microstructure effects.

When neither high-frequency data nor detailed microstructure information is available, practitioners may need to revert to robust but less powerful square-variation-based tests (e.g., MINRV, MEDRV). Nonetheless, the expanding suite of noise-robust jump tests substantially advances the practical and theoretical toolkit for analyzing discontinuities in high-frequency financial processes (Bibinger et al., 2024, Bibinger et al., 2014, Maneesoonthorn et al., 2017, Clinet et al., 2017).