Realized Semivariances: Decomposition & Risk Analytics

• Realized semivariance is a nonparametric estimator that decomposes total return variance into negative (downside) and positive (upside) components for asymmetric volatility analysis.
• It employs high-frequency data along with subsampling and grid shifting to mitigate noise and distinguish between continuous and jump-induced variations.
• Empirical findings demonstrate that downside semivariance significantly improves tail risk forecasting and conditional quantile models in financial risk management.

A realized semivariance is a nonparametric, high-frequency estimator that decomposes the total ex-post variance of asset returns into contributions from negative and positive return increments, denoted as downside and upside semivariance respectively. Unlike standard realized variance, which aggregates squared returns without regard to sign, realized semivariances provide a directional sensitivity that is crucial for modeling asymmetric volatility phenomena and tail risks. Both empirical modeling and large-scale data infrastructures have established the realized semivariance as a key component in financial econometrics, particularly for quantile prediction and risk management (Zikes et al., 2013, Cipollini et al., 23 Feb 2026).

1. Mathematical Definition and Decomposition

Realized semivariance divides total variance measured from high-frequency data into components associated with negative ("downside") or positive ("upside") returns. Given observed (log) prices $p_k$ at regular time intervals, intraday returns are defined as $r_k = \log(p_k/p_{k-1})$ for $k=1,\ldots,m$. Downside and upside realized semivariance over a trading day with $m$ intervals are given by: $$\mathrm{rs}^- = \sum_{k=1}^{m} r_k^2\,\mathbf{1}_{\{r_k<0\}} \qquad \mathrm{rs}^+ = \sum_{k=1}^{m} r_k^2\,\mathbf{1}_{\{r_k>0\}}$$ The unconditional realized variance is recovered as $$\mathrm{rv} = \sum_{k=1}^m r_k^2 = \mathrm{rs}^- + \mathrm{rs}^+$$. This decomposition provides a sign-aware measure of volatility and is essential for detecting the asymmetric impact of return shocks (Zikes et al., 2013, Cipollini et al., 23 Feb 2026).

2. Quadratic Variation, Continuous and Jump Components

In continuous-time modeling, log-price evolves as an Itô semimartingale: $$X_t = X_0 + \int_0^t \mu_s\,ds + \int_0^t \sigma_s\,dW_s + J_t$$ with quadratic variation

$$QV_t = \int_0^t \sigma_s^2\,ds + \sum_{0<s \leq t} (\Delta J_s)^2 \equiv IV_t + JV_t$$

where $IV_t$ is integrated variance and $JV_t$ is jump variation. Realized variance $RV_{t,M}$ consistently estimates $IV_t + JV_t$ as the sampling frequency $M \to \infty$. Semivariances further decompose quadratic variation by the sign of return increments, with the limit

$$RS_{t,M}^- \xrightarrow{p} 0.5 \, IV_t + \sum_{s\in(t-1,t]} \mathbf{1}_{\{\Delta J_s<0\}} (\Delta J_s)^2,$$

providing a partition between negative and positive sources of volatility (Zikes et al., 2013).

3. Estimation Methodologies and Pipeline

High-frequency estimation of realized semivariance requires careful data preparation and noise-mitigation. Implementations such as VOLARE (Cipollini et al., 23 Feb 2026) adopt a multistage pipeline:

• Raw tick ingestion: Acquisition of millisecond-resolved tick data for equities or second-resolved midquotes for futures/FX.
• Outlier cleaning: Local filtering (e.g., Brownlees & Gallo, 2006) flags and interpolates outlier prices to suppress spurious noise.
• Regular-interval sampling: Application of the previous-tick rule on a 1- or 5-minute fixed grid, with non-traded intervals filled forward.
• Return computation: Log-returns computed over regularized price series.
• Aggregation: Basic semivariance computed as above or, for noise mitigation, through "5-subsample" averaging over grid shifts: $$\mathrm{rs}^-_{\mathrm{ss}} = \frac{1}{5} \sum_{s=1}^{5}\sum_{i=1}^{m_s} r^2_{i,s} \mathbf{1}_{\{r_{i,s} < 0\}}$$ Days with insufficient observations are excluded to reduce sampling error (Cipollini et al., 23 Feb 2026).

4. Noise Robustification and Limitations

Microstructure noise complicates high-frequency estimation, potentially distorting the realized semivariance by flipping the sign of small returns and inflating spurious volatility. VOLARE (Cipollini et al., 23 Feb 2026) addresses this with subsampling and grid shifting to average out noise-induced bias and variance. For realized variance, realized kernels (e.g., Parzen-kernel autocovariances) are employed, though not directly for semivariances. Practical considerations include:

• Exclusion of days with fewer than 40 valid intervals
• Awareness of odd-lot filtering and its minimal impact at minute-level
• Potential for jump contamination: semivariance estimators include jump variation by construction and do not distinguish continuous from discontinuous paths.

Sampling frequency is a key parameter: lower intervals reduce microstructure noise but may understate intra-day volatility asymmetries; higher frequencies increase information but amplify noise and sign misclassification (Cipollini et al., 23 Feb 2026).

5. Semivariances in Conditional Quantile Model Frameworks

Semiparametric quantile regression enables the modeling of conditional return and volatility quantiles as functions of realized measures, including semivariances (Zikes et al., 2013). The conditional $\alpha$-quantile of next-day return is expressed as: $$q_\alpha(r_{t+1}\mid\Omega_t) = \beta_0(\alpha) + \boldsymbol\beta_v(\alpha)'\mathbf{v}_{t,M} + \boldsymbol\beta_z(\alpha)'z_t,$$ with $\mathbf{v}_{t,M}$ containing $RV^{1/2}$, $IV^{1/2}$, $JV^{1/2}$, $RS^{-,1/2}$, $RS^{+,1/2}$, etc., and $z_t$ denoting exogenous controls such as implied volatility.

For realized volatility quantile modeling (HARQ framework), the regressors may directly include $\mathrm{rs}^-$ and $\mathrm{rs}^+$. Estimation leverages interior-point quantile regression algorithms and bootstrapped standard errors to account for serial dependence (Zikes et al., 2013).

6. Empirical Findings, Applications, and Comparative Properties

Analysis of both S&P 500 and WTI crude oil futures demonstrates:

• Downside semivariance dominates: $RS^-$ enters strongly and significantly in return quantile regressions, its effect persists even after controlling for implied volatility. Upside semivariance $RS^+$ exhibits negligible and statistically insignificant contributions.
• Jump variation is negligible: Once continuous or semivariance measures are included, jump variation $JV$ does not significantly predict either returns or realized volatility quantiles.
• Implied volatility adds marginal value: Option-implied volatility indices (VIX, OVX) may enter as significant regressors but do not subsume the forecasting power of $RS^-$.
• Forecast accuracy: Incorporating $RS^-$ (and implied volatility) in linear quantile regressions matches or outperforms CAViaR and mixture-ARFIMA models, especially for tail quantiles and medium-horizon realized volatility (Zikes et al., 2013).

In practical settings, realized semivariance supports real-time risk analytics (e.g., Value-at-Risk estimation, volatility forecasting) and underpins asymmetric volatility modeling in high-frequency asset return analysis (Cipollini et al., 23 Feb 2026).

7. Distinctions Versus Other High-Frequency Volatility Estimators

Realized semivariance has several distinguishing properties:

• Directional decomposition: Unlike realized variance or bipower variation, semivariance quantifies asymmetry between negative and positive return-induced volatility.
• Sensitivity to jumps: Includes both continuous sample path and jump-induced returns, in contrast to bipower variation which aims to isolate the continuous component.
• Feature for nonlinear modeling: Provides richer, sign-aware input features for forecasting models (e.g., HAR, MEM).
• Practicality for tail-risk prediction: Downside semivariance is especially informative for applications targeting value-at-risk and lower-tail risk estimation (Cipollini et al., 23 Feb 2026).

Summary table:

Estimator	Jump-Robust	Directional Info	Application Focus
Realized Variance	No	No	Total volatility
Bipower Variation	Yes	No	Continuous component
Semivariance	No	Yes	Downside (tail) risk, asymmetry

Realized semivariances are now standard in empirical asset-pricing and risk management environments, as reflected in large-scale archives such as VOLARE (Cipollini et al., 23 Feb 2026) and leading model validation frameworks (Zikes et al., 2013).