QuickVaR: Fast Estimation of VaR & ES

• QuickVaR is a collection of fast, adaptable risk estimation methods that compute VaR and ES while reducing latency and calibration burdens.
• It incorporates techniques like regime-aware simulation, joint quantile regression, and extreme-value corrections to improve backtesting performance.
• QuickVaR spans both computational algorithms and analytical approaches, including linear-time selection and machine-learning estimators, to suit varied financial applications.

QuickVaR is used in the recent literature as a label for fast Value-at-Risk estimation, but the collected sources suggest that it is not a single standardized model. Instead, it denotes a family of procedures whose common objective is to reduce the latency, variance, or calibration burden of VaR and, in many cases, Expected Shortfall (ES), while preserving acceptable backtesting behavior. These procedures range from regime-aware Gaussian-mixture Monte Carlo and multivariate joint quantile regression to bias-reduced extreme-value methods, jump-aware analytical formulas, conformalized machine-learning estimators, and an expected-linear-time selection algorithm for discrete distributions (Seyfi et al., 2020, Merlo et al., 2021, Agrawal et al., 6 Jul 2026).

1. Conceptual scope and defining conventions

Across the cited works, QuickVaR is anchored in standard quantile-based risk measurement, but not in a single sign convention. In the loss convention used by several financial forecasting papers, VaR and ES are expressed as

$$\mathrm{VaR}_\alpha = -\inf\{x : F_L(x) \ge \alpha\}, \qquad \mathrm{ES}_\alpha = -\mathbb{E}[L \mid L > \mathrm{VaR}_\alpha].$$

In the discrete algorithmic setting, however, the paper on linear-time monetary risk measures adopts a reward convention and defines

$$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$

with the associated right-quantile interpretation for discrete distributions (Seyfi et al., 2020, Agrawal et al., 6 Jul 2026).

This divergence in sign convention is not a minor notational issue. Several papers target financial losses or negative returns directly, whereas the discrete-computation paper treats lower-tail rewards as the risky tail. A recurrent source of confusion is therefore the assumption that all QuickVaR formulations are interchangeable at the formula level. They are not: moving between loss- and reward-based formulations generally requires a sign transformation.

Family	Representative papers	Core mechanism
Regime-aware simulation	(Seyfi et al., 2020)	GMM-based Monte Carlo with stratified sampling
Joint quantile regression	(Merlo et al., 2021, Candila et al., 2020)	Dynamic VaR/ES regressions with dependence modeling
Quantile-scale and compensatory methods	(Liu et al., 2 Mar 2026, Yang, 2021)	Distribution-free scale dynamics or coverage correction
Extreme-tail methods	(Kratz, 2013, Kaibuchi et al., 2021, Martín et al., 2023)	EVT, bias reduction, or informative priors
Jump-aware analytics	(Lin et al., 2023)	Closed-form or semi-analytical jump-risk VaR
Real-time and discrete algorithms	(Wang et al., 2 Feb 2026, Agrawal et al., 6 Jul 2026)	QRF with conformal calibration or Quickselect-style VaR

A further distinction concerns scope. Some QuickVaR variants are end-to-end engines producing rolling forecasts, scenario generation, backtesting, and sometimes portfolio allocation. Others are narrowly computational, replacing sorting or simulation with a faster quantile-evaluation routine.

2. Regime-aware Monte Carlo and mixture simulation

In "Portfolio Risk Measurement Using a Mixture Simulation Approach" (Seyfi et al., 2020), QuickVaR is a fast, practical VaR/ES engine built on Gaussian-mixture-model Monte Carlo. The central object is the mixture density

$$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$

fit on multivariate log returns

$r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$

The stated rationale is that financial returns are non-normal, with heavy tails, skewness, and regime shifts, so a GMM can encode calm and crisis regimes directly through component-specific means, covariances, and weights.

The model is estimated by EM using the standard responsibilities

$$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$

with ridge regularization, covariance shrinkage, floor weights, and model selection by AIC or BIC. The implementation details emphasize rolling windows of 1–3 years, warm-start EM for daily recalibration, and either full or diagonal covariance matrices. The diagonal variant is intended for large-dimensional settings, retaining regime-dependent volatility clustering and conditional co-movement through the shared component index $k$.

Scenario generation is stratified rather than purely multinomial. Each component receives $N_k=\mathrm{round}(\pi_k N)$ simulated draws, and sampled returns are volatility-rescaled through

$$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$

where the data block recommends $\sigma_{Short}\approx 60\text{–}80$ days and $\sigma_{Long}\approx 252$ days. Antithetic pairs and Sobol sequences are listed as variance-reduction options. VaR and ES are then estimated empirically from portfolio losses after compounding over the horizon.

The defining claim of this formulation is that no separate correlation matrix or copula calibration is required: dependence is learned through the component covariances, and crisis periods shift weights toward high-variance, high-correlation components. The empirical summary reports higher KS $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$0-values, lower RMSE relative to Normal and $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$1 alternatives, and generally successful Kupiec and Christoffersen tests for $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$2, while historical simulation and delta-normal frequently failed during turbulent periods (Seyfi et al., 2020).

3. Joint VaR–ES regression and mixed-frequency forecasting

A second major QuickVaR lineage is regression-based and directly targets conditional quantiles and shortfall. In "Forecasting VaR and ES using a joint quantile regression and implications in portfolio allocation" (Merlo et al., 2021), the state variable is the vector of asset returns $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$3, whose conditional quantiles $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$4 evolve through CAViaR-type autoregressions. The paper considers SAV, AS, and IG news-impact specifications and combines them with either a multiplicative ES link,

$$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$5

or an autoregressive ES gap. Cross-asset dependence is introduced by a time-varying multivariate asymmetric Laplace distribution,

$$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$6

with $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$7 as the correlation matrix inside $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$8.

This construction is designed to exploit the joint elicitability of VaR and ES through the MAL likelihood and the associated multivariate loss $$\operatorname{VaR}_\alpha(X) = \inf\{ t \in \mathbb{R} : F_X(t) \ge \alpha \},$$9. Estimation proceeds by EM using the MAL location-scale mixture representation, with univariate CAViaR initialization, empirical-correlation initialization for $$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$0, and multiple random starts. In the reported weekly-index application, the multivariate CAViaR-AS with multiplicative ES passed coverage tests consistently and yielded lower average losses than univariate benchmarks. The portfolio-allocation extension relies on the fact that linear combinations of MAL vectors are AL, giving tractable portfolio VaR and ES through explicit transformed parameters (Merlo et al., 2021).

A related but distinct mixed-frequency implementation appears in "Mixed-frequency quantile regressions to forecast Value-at-Risk and Expected Shortfall" (Candila et al., 2020). There, daily returns $$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$1 are decomposed into low-frequency and high-frequency components, with the low-frequency term entering through a MIDAS Beta-weighted aggregator

$$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$2

The conditional quantile is then

$$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$3

ES is linked to VaR through an ALD-based specification,

$$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$4

In the reported crude-oil and gasoline application, MF-QR-X passed VaR UC, CC, and DQ tests and ES UC and CC tests, outperforming a range of GARCH, GJR, GARCH-MIDAS, CAViaR, and historical-simulation competitors (Candila et al., 2020).

4. Quantile-scale dynamics, moment updating, and compensatory correction

A third group of methods accelerates VaR estimation by avoiding full distributional modeling. "Quantile-based modeling of scale dynamics in financial returns for Value-at-Risk and Expected Shortfall forecasting" (Liu et al., 2 Mar 2026) defines conditional scale as the difference between two conditional quantiles,

$$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$5

with $$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$6 and $$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$7. After estimating the conditional median by QAR(1), the method fits restricted global CAViaR equations for the lower and upper quantiles, computes normalized residuals, and estimates VaR through

$$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$8

ES is approximated by averaging conditional quantiles below $$f(r) = \sum_{k=1}^K \pi_k \,\mathcal{N}(r; \mu_k, \Sigma_k), \qquad \sum_{k=1}^K \pi_k = 1,\ \pi_k \ge 0,$$9. The data block reports that QbSD-gAS ranked strongly in model confidence sets across major equity indices, particularly at $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$0 and $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$1, and often outperformed established GARCH and joint VaR–ES conditional-quantile methods (Liu et al., 2 Mar 2026).

A faster corrective strategy is provided in "Compensatory model for quantile estimation and application to VaR" (Yang, 2021). Here the baseline quantile $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$2 is not replaced but adjusted by observed coverage error:

$r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$3

Equivalently, adjusted VaR is

$r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$4

The paper proves bounds showing that the empirical coverage error contracts at rate $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$5 under its assumptions. In the reported S&P 500 example, increasing $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$6 from $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$7 to $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$8 moved empirical coverage close to target and materially improved Kupiec and Christoffersen $r_{t,j} = \ln(S_{t,j}/S_{t-1,j}).$9-values (Yang, 2021).

A more classical fast-update route appears in the EWMA framework with time-varying variance, skewness, and kurtosis (Gabrielsen et al., 2012). There, variance, skewness, and kurtosis are updated recursively by separate exponential decays, and VaR is obtained by Cornish-Fisher expansion. This approach is computationally light because each update is $$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$0, requiring only the current state and new return.

5. Extreme tails, heavy tails, and jump-aware formulations

Several QuickVaR variants are motivated by failure of Gaussian or CLT-based approximations in the tail. "There is a VaR beyond usual approximations" (Kratz, 2013) introduces Normex for aggregated heavy-tailed risks. The method decomposes the sum $$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$1 into a trimmed bulk and an extremes block, approximates the conditional bulk by a normal law, and treats the largest $$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$2 order statistics explicitly. The rule

$$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$3

with $$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$4 is recommended to ensure finiteness of the relevant trimmed moments. In the reported Pareto simulations, Normex produced near-exact regulatory quantiles, with errors often below $$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$5, whereas CLT systematically underestimated high quantiles and the maxima approximation was less accurate (Kratz, 2013).

For conditional extreme VaR in financial time series, "GARCH-UGH: A bias-reduced approach for dynamic extreme Value-at-Risk estimation in financial time series" (Kaibuchi et al., 2021) filters losses through AR(1)-GARCH(1,1),

$$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$6

then estimates the innovation tail quantile using the Unbiased Gomes–de Haan estimator on standardized residuals. The one-step-ahead VaR is

$$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$7

The reported backtesting summary attributes fewer Kupiec and Christoffersen failures to GARCH-UGH than to GARCH-EVT or unfiltered UGH, together with lower sensitivity to the tail fraction $$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$8 (Kaibuchi et al., 2021).

A Bayesian EVT alternative appears in "New Bayesian method for estimation of Value at Risk and Conditional Value at Risk" (Martín et al., 2023). Exceedances above a threshold $$\gamma_{ik} = \frac{\pi_k \,\mathcal{N}(r_i;\mu_k,\Sigma_k)} {\sum_{\ell=1}^K \pi_\ell \,\mathcal{N}(r_i;\mu_\ell,\Sigma_\ell)},$$9 are modeled by a GPD with informative priors derived from the baseline distribution. For exceedances $k$0, the paper uses

$k$1

with mapping back to the baseline through $k$2. The data block reports that the informative-prior MH approach was more accurate and precise than exceedance-only MH, particularly at small sample sizes (Martín et al., 2023).

Jump-aware QuickVaR is represented by "Estimation of VaR with jump process: application in corn and soybean markets" (Lin et al., 2023). The model starts from a geometric Lévy price process with Brownian and jump components, then uses change-point detection to identify and remove cumulative jumps from prices before estimating $k$3, $k$4, and $k$5. In analytically convenient cases, expected VaR reduces to a lognormal-type expression,

$k$6

In the reported 99% daily VaR series for corn and soybean markets, the jump-filtered formulation generally produced larger VaR than the version without jumps, and the Dynamic Quantile test $k$7-values exceeded the 1% significance threshold for all markets (Lin et al., 2023).

6. Discrete linear-time computation and real-time machine-learning QuickVaR

The most literal use of the name occurs in "Computing Monetary Risk Measures in Linear Time" (Agrawal et al., 6 Jul 2026). There, QuickVaR is an algorithm for computing the VaR of a discrete random variable in expected linear time by adapting randomized Quickselect. Instead of sorting the full support, it chooses a pivot $k$8, partitions the domain into $k$9, $N_k=\mathrm{round}(\pi_k N)$0, and $N_k=\mathrm{round}(\pi_k N)$1, computes the probability masses $N_k=\mathrm{round}(\pi_k N)$2 and $N_k=\mathrm{round}(\pi_k N)$3, and recurses according to whether $N_k=\mathrm{round}(\pi_k N)$4, $N_k=\mathrm{round}(\pi_k N)$5, or $N_k=\mathrm{round}(\pi_k N)$6. The expected complexity is $N_k=\mathrm{round}(\pi_k N)$7, memory is $N_k=\mathrm{round}(\pi_k N)$8 in place, and the same paper builds QuickDivergence on top of QuickVaR to compute $N_k=\mathrm{round}(\pi_k N)$9-divergence risk measures, including CVaR and TVaR, through a polymatroid formulation. The numerical summary reports an order-of-magnitude speedup for large domains, and the paper provides a Julia implementation in RiskMeasures.jl (Agrawal et al., 6 Jul 2026).

A separate real-time direction is developed in "Reliable Real-Time Value at Risk Estimation via Quantile Regression Forest with Conformal Calibration" (Wang et al., 2 Feb 2026). This method uses the offline-simulation-online-estimation framework: conditional losses are generated offline, a quantile regression forest learns $$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$0, and online VaR is obtained by fast forest scoring,

$$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$1

Reliability is then restored by conformal calibration:

$$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$2

Under exchangeability, the calibrated estimator satisfies

$$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$3

The reported experiments show that conformalized QRF achieves mean coverage rates close to target and improves pinball loss relative to uncalibrated QRF, especially at extreme confidence levels (Wang et al., 2 Feb 2026).

Taken together, these two papers show that QuickVaR can mean either computational acceleration of quantile extraction itself or real-time conditional VaR estimation with explicit coverage control.

7. Backtesting standards, practical choices, and recurrent limitations

Despite their methodological differences, QuickVaR variants converge on a common validation toolkit. Kupiec unconditional coverage, Christoffersen independence or conditional coverage, Dynamic Quantile tests, and ES-specific procedures such as Du–Escanciano or Acerbi–Szekely tests recur across the literature (Seyfi et al., 2020, Merlo et al., 2021, Candila et al., 2020, Lin et al., 2023). This commonality is significant because the core claim of being “quick” is not treated as sufficient; the faster method is expected to retain or improve regulatory and statistical acceptability.

Several practical recommendations are repeated. Rolling estimation windows are standard; warm starts, diagonal or shrinkage covariances, and low-dimensional factorizations are used when dimension is large; tail procedures require careful threshold or $$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$4-selection; and very short volatility or calibration windows are often reported as unstable. In end-to-end forecasting pipelines, daily or weekly re-estimation is common, whereas the linear-time discrete algorithm and the EWMA-style moment recursion are designed for essentially immediate recomputation.

The literature also identifies recurrent limitations. EM-based mixtures can overfit when $$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$5 is too large or windows are too short; MAL dependence estimation can become noisy at higher dimension; EVT procedures remain sensitive to thresholding and tail sparsity; compensatory quantile correction can become overly conservative if $$r^{adj}_{j} = r_j \times \frac{\sigma_{Short,j}}{\sigma_{Long,j}},$$6 is too large; conformal validity relies on exchangeability, which time series only approximate; and reward- versus loss-based conventions can create implementation errors if formulas are transplanted without sign adjustments (Yang, 2021, Wang et al., 2 Feb 2026, Agrawal et al., 6 Jul 2026).

The overall pattern suggests that QuickVaR is best understood as an architectural objective—fast, backtest-conscious quantile risk estimation—rather than as a single estimator. In some papers that objective is achieved by regime-aware simulation, in others by direct conditional-quantile modeling, tail asymptotics, analytical jump filtering, or selection-style algorithms. The unifying criterion is not model form but the attempt to compute or forecast VaR, and often ES, with lower latency and stronger empirical calibration than slower or more rigid baselines.