VAR-Scaling in Quant Finance

• VAR-Scaling is a collection of methodologies that extend daily Value-at-Risk estimates to longer horizons by incorporating scaling laws and adjustments for fat tails and volatility clustering.
• The approach leverages the square-root-of-time rule and its extensions—including simulation and GARCH models—to more accurately capture risk under non-iid conditions.
• VAR-Scaling is critical for both regulatory risk management and advanced applications in high-dimensional time series and image generative modeling, enhancing computational efficiency and accuracy.

Value-at-Risk (VaR) scaling refers to a collection of methodologies for transforming short-horizon risk measures—typically expressed as VaR computed over one day—to longer holding periods, such as ten days or a year. The broad aim is to estimate the potential loss at a given confidence level over extended time intervals, either for internal risk management or regulatory purposes. VAR-scaling has deep roots in quantitative finance, rooted in assumptions about the temporal aggregation of returns, and extends to contemporary modeling challenges—including high-dimensional VAR time series estimation and scale-adaptive generative inference in visual autoregressive (VAR) models.

1. Mathematical Foundations of VaR Scaling

In the classical setting, VaR quantifies the maximum loss not exceeded with a specified probability (e.g., 99%) over a given time horizon. Consider a stationary portfolio where returns follow a Gaussian process with mean $\mu$ and variance $\sigma^2$ per unit time $t_0$. The foundational result is the “square-root-of-time” rule:

$$\operatorname{VaR}_{1-p}(t_0) = \mu t_0 + z_{1-p}\sigma\sqrt{t_0}$$

where $z_{1-p}$ is the standard normal quantile. Extending to a horizon $T=k t_0$, the variance accumulates linearly and the scaling law becomes:

$$\operatorname{VaR}_{1-p}(T) = \operatorname{VaR}_{1-p}(t_0) \sqrt{\frac{T}{t_0}}$$

For $t_0 = 1$ day, this is prevalent in banking regulations, notably under Basel III/IV, where 1-day VaR values are scaled by $\sqrt{T}$ to obtain VaRs for longer periods (Kuhlmann, 2022).

Extensions beyond Gaussianity—leveraging distributions with heavier tails (e.g., Student’s $t$, Variance-Gamma)—require convolution-based scaling or explicit calculation of far-tail quantiles. Under extreme value theory (EVT), block maxima feature $1/\alpha$ scaling, with $\alpha$ being the tail index (Cotter, 2011):

$$\operatorname{VaR}_{h, p} = h^{1/\alpha} \operatorname{VaR}_{1, p}$$

The scaling relation is contingent upon fundamental assumptions: independence, identical distribution, stationarity, and finite second moments. Departure from these assumptions, especially the appearance of fat tails, volatility clustering, or autocorrelation, necessitates model-based horizon aggregation (e.g., via GARCH) or simulation-based multiscaling approaches (Spadafora et al., 2014, 2002.04164).

2. Statistical Assumptions and Model-Based Extensions

The naive $\sqrt{T}$ rule constitutes a thin-tailed, iid benchmark. Empirical financial return series are often non-Gaussian, exhibiting pronounced leptokurtosis, skewness, and time-varying volatility. Violations of the iid assumption can have substantial quantitative impact:

• Serial dependence: Volatility clustering refutes the independence assumption, leading to underestimated risk if scaling is naively applied.
• Fat tails: Leptokurtic returns (e.g., low degrees-of-freedom $t$-distributed residuals) inflate VaR relative to the $\sqrt{T}$ rule, especially at high confidence levels (Spadafora et al., 2014).
• Time-varying volatility: GARCH(1,1) models forecast conditional variances over horizon $T$ directly, avoiding misspecification arising from static scaling.

Kuhlmann’s empirical analysis (Kuhlmann, 2022) demonstrates that direct multi-day VaR estimation outperforms scaled VaR—particularly under volatility clustering or regime shifts. Where data scarcity forces reliance on scaling, alternative exponents ("fractal scaling") or fat-tail adjustments may partially correct for under- or overestimation.

Parametric extensions include:

• Student’s $t$ scaling: $\mu + t_{1-p, \nu}\sigma$ using fitted tail parameters.
• Historical simulation: Nonparametric quantiles computed directly on multi-day P&L, often sidestepping scaling altogether.
• Multi-horizon GARCH: Forecasting conditional volatility at horizon $T$ rather than through aggregation (Kuhlmann, 2022, Spadafora et al., 2014).

3. Data Models, Standardization, and Regulatory Practice

Historical VaR and ES (Expected Shortfall) computations depend critically on scaling transformations applied to past data—termed “Data Models” (Kenyon et al., 2014). Three principal families are codified:

Scaling Family	Mathematical Formulation	Volatility Hypothesis
Absolute-difference	$\Delta P_{t_i}^{\rm abs} = P_{t_i} - P_{t_{i-m}}$	Volatility independent of level
Relative	$r_{t_i} = (P_{t_i} - P_{t_{i-m}})/P_{t_{i-m}}$	Scales linearly with level
Level-relative	$r_{t_i}^{(f)} = f(l_i, l_{\rm now})r_{t_i}$	Parametric dependence

Custom scaling functions $f_{\rm lin}$, $f_{\rm quad}$, etc., ensure volatility is correctly propagated in line with observed tail behavior and current market levels. Standardization across institutions—via published lookup tables—eliminates subjective variability and improves interbank consistency in capital calculations.

Regulatory guidance increasingly emphasizes demonstrable empirical validity of scaling assumptions, routine backtesting, and documentation of the data transformation applied (Kenyon et al., 2014, Kuhlmann, 2022).

4. Multiscaling, Statistical Testing, and Advanced Risk Estimation

Scaling exponents may themselves vary with moment order, motivating multiscaling: $H_q = A + Bq$ (generalized Hurst exponent). Brandi and Di Matteo pioneer methodologies—RNSGHE and Monte Carlo MSVaR simulation—for robust estimation and testing of multiscaling properties, reducing bias in annual VaR forecasts relative to traditional “$\sqrt{T}$” scaling (2002.04164). Structure functions and autocorrelation segmented regression are key tools to calibrate the relevant time windows; t-tests and F-tests on $H(q_i,q_j)$ increments enable discrimination between true and spurious scaling.

Monte Carlo simulations employing multifractal random walks (MRW), calibrated to empirical scaling exponents, yield aggregate risk estimates matching observed annual VaR in the majority of tested equities (2002.04164).

5. Computational Scaling in VAR Time Series and Generative Models

VAR-scaling is also central in high-dimensional time series and generation tasks. Standard VAR($p$) models exhibit quadratic (or cubic) scaling in parameter count and computational cost as the number of series $N$ grows (Hu et al., 2022). Neighborhood-VAR (NVAR) and panel VAR frameworks mitigate this by exploiting sparse neighborhood or shared low-rank structure, reducing parameterization from $O(N^2p)$ to $O(N\bar\tau p)$ or through low-rank plus sparse decomposition, with ADMM optimization and theoretical consistency results (Xu et al., 18 Sep 2025).

For visual autoregressive generative models, VAR-Scaling refers to algorithms that optimize compositional inference across scales—either at test/inference-time (see TTS-VAR (Chen et al., 24 Jul 2025), VAR-Scaling (Tang et al., 12 Jan 2026), FastVAR (Guo et al., 30 Mar 2025)) or in the underlying architecture (SRDD as discrete diffusion (Kumar et al., 26 Sep 2025)). Complexity growth with image resolution is tamed via batch-size scheduling, hybrid sampling, clustering-based diversity, cached token pruning, or leveraging Markovian masking to match diffusion sampling strategies.

Empirical results demonstrate that adaptive scaling at coarse scales yields substantial gains in sample fidelity, computational efficiency, and diversity (GenEval, FID, IS metrics). FastVAR achieves up to $2.7\times$ wall-clock speedups for high-res generation on commodity GPUs with negligible performance drop (Guo et al., 30 Mar 2025).

6. Practical Implications, Implementation Strategies, and Recommendations

Across domains, VAR-scaling impacts capital requirements, margin computations, and risk provisioning. In finance, erroneous scaling—especially over longer horizons—may precipitate regulatory arbitrage, excessive capital reserves, or hidden shortfalls. Best practice involves:

• Direct multi-day VaR estimation wherever feasible, especially for assets exhibiting time-dependent volatility.
• Empirical validation of scaling regime via goodness-of-fit, backtesting, and tests for independence and coverage.
• Adoption of standardized data models and lookup tables to reduce systemic variability, as advocated in recent regulatory proposals (Kenyon et al., 2014).
• In generative modeling, descending batch schedules, density-adaptive sampling, and iterative refinement maximize performance for realistic scaling of image content.

Banks, asset managers, and risk system developers are advised to critically evaluate the theoretical premises underpinning their chosen scaling rule, to routinely backtest both scaled and direct VaR estimates, and to document all transformations, parameter inference, and horizon aggregation steps. Where data scarcity or computational constraints force reliance on scaling, parametric corrections or nonparametric simulation should mitigate known biases and model risks.

7. Controversies, Limitations, and Evolving Practice

The continuing centrality of $\sqrt{T}$ scaling in regulation and practice persists despite strong evidence for its empirical failure under heavy-tailed, autocorrelated, or heteroskedastic return series (Kuhlmann, 2022, Spadafora et al., 2014). Recent research advocates for model-based and simulation alternatives, subject to empirical verification and quantification of residual biases.

In machine learning and generative modeling, scaling with respect to resolution and token count remains an active area—optimizing computational cost versus sample diversity and fidelity via adaptive inference-time protocols and architectural modifications (Chen et al., 24 Jul 2025, Tang et al., 12 Jan 2026, Guo et al., 30 Mar 2025, Kumar et al., 26 Sep 2025).

This suggests that the evolution of VAR-scaling is ongoing, with empirical methodologies, probabilistic modeling, and computational techniques converging towards more robust, context-aware, and standardized estimation practices across diverse application domains.