AQAI QuantAI research lab for systematic strategies

Automated analysis

This analysis was drafted by our research engine and has not been checked by a human editor. It may contain errors. It separates the paper’s own results from our tests, and any figures called ours come from our own backtest.

Our automated analysisOur backtest

A one-parameter volatility estimator for sparse contamination

At 5% spikes and lambda 0.9, density-power reads 2.8810 against a truth of 3; Hölder reads 3.0034

2026-09-08 · 9 min read · US equities and ETFs

Reviewing: Robustified Gaussian quasi-likelihood inference for volatility · Shoichi Eguchi and H. Masuda · Read it on openalex

Our backtest of this idea

Our automated quick test, not the paper's

Density-Power Robust Intraday Volatility Targeting for Liquid US Equities and ETFs

Backtest period 2022-01-03 to 2025-12-31 · hypothetical, net of modelled costs

Why these figures are not the paper's (3)

The paper reports no results of its own

This is a theoretical paper — derivations and proofs, with no measurement on market data. The backtest below is a strategy we built from its idea, not a test of anything the authors claimed.

Run on a different market than the paper

The paper trades high-frequency; we have no price data for it, so the strategy below runs on US equities and ETFs instead. The judge did not state why the mechanism survives the swap.

The paper's own figures describe its universe and do not carry over to ours.

Our own audit found this run does not follow the paper faithfully (10)

  • Equation 2.7 observed-increment decomposition (invalidates: Direct application of the paper's finite-activity contamination asymptotics to the empirical minute-price process)
  • Equation 3.14 scaled-estimator limit (invalidates: Theorem 3.5 √n asymptotic mixed-normal guarantee for this empirical backtest)
  • Equation 3.22 Fisher-information matrix (invalidates: The clean-limit covariance I(θ_0)^{-1} as a guaranteed covariance for the trading adaptation)
  • Equation 3.23 fixed-λ asymptotic mixed normality (invalidates: Fixed-λ asymptotic mixed-normal inference and equation 4.1 confidence-interval calibration for the empirical backtest)

6 further finding(s) are described in the note.

These are our findings about our own implementation, not criticisms of the paper. Read the figures below as a description of what we ran.

Jan 2022Total 15.8%Dec 2025
Sharpe
0.48
Total Return
15.8%
Max Drawdown
-20.6%
CAGR
3.8%
Volatility
8.3%
Beta vs SPY
0.17
Trades
24,060

A trader gets a closed-form objective and one tuning value, yet that value changes with the contamination. The authors choose 0.2 for spikes and 0.2 again for jumps sized N(0,3). With Gamma(1,1) jump sizes, the best setting moves to 0.7.

One disclosure governs the final section. Eguchi and Masuda built their estimator for high-frequency sampling, specifically tick-level increments of a continuous semimartingale. We have no tick data. Our exercise uses 1-minute US equity and ETF bars, so it belongs beside the paper rather than serving as a replication. The paper reports no empirical work of any kind.

The target is a volatility regression. The observed process is a d-dimensional continuous Itô semimartingale Y, with diffusion coefficient σ(X, θ) determined by a covariate process X and the parameter. Volatility therefore depends on more than Y alone. Observations arrive at t_j = jT/n, with T fixed as n grows, while the drift remains a nuisance.

Gaussian quasi-likelihood fits θ as though every increment were normal, with mean zero and variance h·S(X_{t_{j-1}},θ). The criterion sums log φ(ΔY; 0, hS_{j-1}(θ)) over the increments. Its score contains the unbounded term (ΔY)²/h. Against ordinary increments of size O(√h), a single jump of size O(1) can pull the estimate almost anywhere.

Threshold filters are the usual response. An implementer discards an increment whenever |ΔY| > (Δt)^ρ C, choosing C and ρ, or uses a global version of that rule. Eguchi and Masuda instead multiply each summand by its own Gaussian density raised to λ. Once an increment standardizes to a huge value, its weight falls like exp(−λ|z|²/2) and its contribution nearly disappears. This requires neither a threshold nor a preliminary spot-volatility estimator.

The paper develops two forms. Its density-power quasi-likelihood (eq. 3.6) restores a compensation constant, K_{λ,d} = (2π)^{−dλ/2}/(λ+1)^{1+d/2}, preserving an approximate martingale estimating equation. The Hölder-normalized form (eq. 3.12) comes from the γ-divergence direction and divides by the Hölder normalizer instead of deducting a constant.

Theorem 3.5 establishes asymptotic mixed normality at rate √n for every fixed λ. A shrinking λ is covered as well. When λ_n → 0 slowly enough to satisfy √n h^κ/λ_n → 0 with κ > 1/2, the limiting covariance reduces to I(θ_0)^{-1}, matching the efficient contamination-free case.

Every result is Monte Carlo. The authors use the yuima R package, running 1000 independent sample paths for each configuration, with n = 1000 and 5000 and T = 1. Their first model is a time-inhomogeneous Wiener regression, θ_0 = (−2, 3, 0), driven by deterministic trigonometric covariates. The second is a jump-diffusion with θ_0 = (2, 3).

Contamination takes two forms. For spikes, Bernoulli(p) draws determine isolated errors with N(0,σ²) magnitude. The settings are p = 0.01 and 0.05, paired with σ² = 1 and 3. For jumps, compound Poisson intensity is q = 0.01n and 0.05n, while sizes follow N(0,3) or Gamma(1,1). The paper's spike noise is measurement error at isolated timestamps, of the kind cited from echo-planar diffusion tensor imaging. Market microstructure noise behaves differently. The paper contains no empirical application and reports no economic performance.

The simulation failure of the naive estimator is severe. Under 1% spike contamination, with σ² = 1 and n = 5000, the naive GQMLE produces mean estimates of −0.1243 and 0.1429 for true values of −2 and 3. Standard deviations are near 0.36. In the same cell, density-power at λ = 0.5 gives −1.9916 and 2.9920, with standard deviations of 0.0361 and 0.0356.

The jump-diffusion case is harsher. At q = 0.05n and n = 5000, the naive estimator reaches the optimizer's upper bound in every run: mean 10.0000 and 10.0000, standard deviation 0.0000 across all 1000 replications, versus a truth of (2, 3).

The theorem draws a narrow boundary

Assumption 2.2 allows finitely many jumps on [0, T]. Under Assumption 2.3, spikes may occur at only finitely many observation times, uniformly in n. These conditions are explicit, and they define the theorem's reach. Additive noise at every timestamp falls outside that reach. My reading is that intraday quote data violates the condition directly because contamination touches every observation rather than a small set. An intraday desk needs protection from noise in every increment. This theorem promises something narrower.

The limit also treats the two sources asymmetrically. Its random covariance retains jumps in the covariate process X*, whereas spikes vanish. Jumps in the explanatory process remain; measurement error washes out.

Theorem 3.5 further assumes Γ_0(λ) is a.s. positive definite. In practice, that can only be checked with a plug-in estimate. The paper supplies the construction: consistent estimators of Γ_0(λ), Σ_0(λ) and I(θ_0) follow from the uniform convergence in (3.25). Its simulations calculate Γ_n and Σ_n within the studentization (4.1).

Σ_0(λ) raises the weaker point. Positive definiteness carries no guarantee for fixed λ > 0, as the paper states directly. The offered argument uses continuity at zero and therefore establishes the property for λ small enough.

The authors examine a logarithmic Hölder-divergence variant, H_log,n, and reject it. Heterogeneity makes the normalizing integral depend on j. As a result, the log form does not yield an approximate martingale estimating function. Their conclusion is to use H_n(θ;λ) rather than H_log,n(θ;λ).

How much can one tuning value carry?

The abstract says the simulations "highlight the estimators' insensitivity to the choice of the tuning parameter." Remark 4.1 gives a less reassuring account. The choice "is important, and the optimal value may depend on the configuration considered." The authors then write: "In the literature on density-power divergence-based estimators, values around 0.2 are commonly used as a tuning parameter, providing a good balance between efficiency and robustness."

Their coverage tables complicate that defence. At λ = 0.2, σ² = 1, p = 0.01 and n = 5000, density-power coverage is 0.92, 0.92 and 0.91 across the three coefficients. Hölder coverage in the same cell is 0.95, 0.94 and 0.95. The suggested tuning value is precisely where the estimators separate.

Higher λ brings more dispersion. With σ² = 1, p = 0.01 and n = 5000 under spike contamination, the standard deviation of density-power θ̂1 rises from 0.0305 at λ = 0.1 to 0.0361 at λ = 0.5, then 0.0456 at λ = 0.9. The taper has a cost, though a limited one.

Bias is more damaging and less even. Under 5% spikes, σ² = 1 and n = 5000, density-power at λ = 0.9 estimates θ̂2 at 2.8810 with standard deviation 0.0461. By our own arithmetic from the paper's reported mean and s.d., that miss is roughly 2.6 times the estimator's own dispersion. In the identical cell, Hölder returns 3.0034 with standard deviation 0.0455.

The paper says both estimators "have similar tendencies" in λ. That warning limits what one cell can establish. I would therefore avoid claiming general tuning insensitivity for the Hölder estimator.

Efficiency and calibration pull in opposite directions. Sending λ_n → 0 recovers the clean-data limit. Yet with λ = 0.1, σ² = 1, p = 0.01 and n = 5000, empirical 95% density-power coverage is 0.88, 0.83 and 0.89 across the three coefficients. For λ ≥ 0.2 in the same table, its range improves to 0.91 to 0.94. Hölder records 0.92, 0.88 and 0.94 at λ = 0.1, then 0.93 to 0.96 for λ ≥ 0.2. Theory favors a small λ, while the variance approximation performs worst at 0.1. Both estimators undercover there.

No simulation comparison against the threshold estimators appears anywhere in the tables.

The introduction sets the method against the global jump filter of Inatsugu and Yoshida and the local-threshold work of Shimizu and Yoshida. The tables contain only the naive GQMLE, density-power GQMLE and Hölder GQMLE. The text makes a fair argument about calibration burden: the global-threshold procedure needs several user choices, among them a preliminary spot-volatility estimate and a threshold sequence. The paper runs no accuracy contest.

We would ship Hölder

Under heavy contamination, Hölder changes less with λ. At λ = 0.9, 5% spikes, σ² = 1 and n = 5000, it returns 3.0034 against density-power's 2.8810. Calibration is stronger through the useful range as well. In the σ² = 1, p = 0.01 and n = 5000 tables, Hölder coverage spans 0.93 to 0.96 for λ ≥ 0.2, compared with 0.91 to 0.94 for density-power.

It is also cheaper. At λ = 0.5 and n = 5000, mean calculation time is 0.3556 s for Hölder and 0.6450 s for density-power. The naive GQMLE takes 0.2651 s.

Residual clustering belongs in a separate discussion. The authors present it as a practical step rather than part of the theory. They choose K by inspecting where the number of flagged observations jumps. On the spike data, with λ = 0.2 and n = 5000, 4-means identifies 29 of 39 noise points correctly. On jump data with q = 0.01 and roughly 50 expected jumps, 6-means flags 38 elements. We would keep that out of a control loop.

Our 1-minute US bar adaptation

As disclosed above, we built adjacent work rather than the paper's estimator. Our universe contains the top 30 US stocks and ETFs by trailing dollar volume, excluding ADRs, using 1-minute regular-session bars.

Each session, we fit the fixed-λ density-power quasi-likelihood at λ = 0.2. The estimation window contains the latest 1,000 within-session increments and never crosses an overnight boundary. Variance follows a log-linear design with three columns: an intercept, the lagged squared scaled increment, and the prior-session variance rate. The fitted model forecasts next-session annualized volatility. We size an equal-weight long book with clip(0.10 / forecast volatility, 0, 1), then execute at the next session's first 1-minute close.

From 2022-01-03 through 2025-12-31, the book returned 15.81% in total. Its Sharpe was 0.48, with maximum drawdown of -20.58% and realized volatility of 8.33% (Sortino 0.66, Calmar 0.18). A 0.48 Sharpe paired with a -20.58% drawdown is weak for a partially de-risked long equity book. The sizing cap drives that outcome more than the estimator does. Since the paper is theoretical and gives no measured performance, our figures cannot be compared with any result of theirs.

Our design largely removes the estimator from the outcome. The multiplier stops at 1, and each instrument uses a 10% volatility reference. Forecasts for liquid megacaps usually sit well above 10%, leaving the overlay below its equal-weight base for much of the sample. Return, volatility and drawdown mostly reflect the fraction of a long equity book switched on. The variance estimate has little influence.

We did not run the cleaner construction. Such a run would send the density-power forecast, naive GQMLE forecast and prior-session realized variance through the same sizing rule, then report the differences.

Our implementation has two weaknesses. Cost treatment is inconsistent: the rule applies a 5 bps commission plus 1 bp slippage per side, alongside a $0.004 per share configuration with zero slippage. The yearly liquidity ranking also needs verification as point-in-time. Selection could enter through that step.

The ranking of volatility estimators changes with the scoring criterion. Our note on EGARCH versus plain GARCH on NEPSE reached the same conclusion. In this paper, disagreement appears across λ rather than across loss functions.

Two results would change my reading. A comparison with the global jump filter would show what accuracy, if any, is surrendered for the lighter calibration burden. Alternatively, Assumption 2.3 could be extended to admit noise at every observation. Until one of those appears, this remains a clean single-parameter estimator for sparse contamination, while intraday equity prices carry contamination throughout the sample.

How our backtest worked

The steps the code we ran actually executed, from its strategy card. Ours, not the paper's — it is one automated implementation of the idea, not the authors' own.

For each calendar year:
    Select the top 30 non-ADR US stocks and ETFs by trailing dollar volume.

For each instrument and session t:
    1. Form log-return increments only between consecutive valid 1-minute
       closes within the same regular session; never bridge overnight or gaps.
    2. Scale each increment as y_j = ΔY_j / sqrt(h), where T = 1,
       n = 1000, and h = T/n = 0.001.
    3. Using the latest 1,000 valid increments available through session t,
       construct x_{j-1} = [1, log(floor + y_{j-1}²),
       log(floor + prior-session realized-variance rate)].
    4. Require a full-rank three-column covariate design. Otherwise skip the fit.
    5. Fit θ within its bounded domain by maximizing the complete fixed-λ
       density-power Gaussian quasi-likelihood with λ = 0.2. Recompute the
       Gaussian-power weight for every candidate θ; do not threshold returns.
    6. Obtain the next-session conditional variance-rate forecast and convert it:
       integrated variance = m_forecast × h × forecast variance rate
       annual volatility = sqrt(252 × integrated variance).
    7. Start from an equal-weight long-only allocation and set the instrument
       multiplier to clip(10% / forecast annual volatility, 0, 1).
       Apply the 10% position cap and portfolio leverage constraints.
    8. Freeze the signal after the current session close and execute at the next
       regular session's first available 1-minute close. Hold until the following
       session's first-close rebalance.
    9. If a required forecast, analytical input, or real execution close is
       unavailable, omit invalid increments or skip the affected rebalance/order;
       never synthesize prices or volume.

Conventional Gaussian quasi-likelihood and prior-session realized volatility are specified as forecast benchmarks, while extreme sessions are analysis-only slices based on an expanding 95th-percentile standardized-increment threshold.