Factorized Tail Volatility Model

• Factorized Tail Volatility Model is a stochastic volatility framework that encapsulates heavy-tailed distributions and flexible extremal dependence via a factorization structure.
• It models latent log-volatility with a factorized sum, enabling precise estimation of tail quantiles and addressing clustering of extremes and tail risk heterogeneity.
• FTVM integrates advanced techniques such as Breiman-type theorems and EoT extensions, demonstrating superior performance in simulations and option pricing applications.

The Factorized Tail Volatility Model (FTVM) is a family of stochastic volatility models designed to capture both heavy-tailed marginal distributions and flexible extremal dependence properties in time series and high-dimensional data. Through a factorization structure applied to volatility and/or tail quantiles, FTVM accommodates empirical features such as clustering of extremes and heterogeneity in tail risk, addressing key limitations of classical stochastic volatility and GARCH frameworks.

1. Formal Model Definition

FTVM is defined on a discrete-time probability space supporting three independent i.i.d. sequences: $(Z_t)$ (innovations), $(\eta_t)$ (log-volatility shocks), and an initial $\sigma$-field for stationarity. The observed process takes the form

$R_t = \sigma_t Z_t, \qquad t \in \mathbb{Z},$

with latent log-volatility given by a (possibly infinite) factorized sum: $$\log\sigma_t = \sum_{i=0}^\infty \alpha_i\,\eta_{t-i},\qquad \alpha_i \in [0,1], \quad \max_i \alpha_i=1,\quad \alpha_i=O(i^{-\theta}),\,\theta>1.$$ The innovations $Z_t$ are i.i.d., symmetric or tail-balanced, satisfying integrability $E|Z_0|^{1+\delta}<\infty$ for some $\delta>0$. The log-volatility shocks $(\eta_t)$ are i.i.d. with $E(\eta_0^2)<\infty$ and subexponential upper tail: $$\Pr\{\eta_0>z\} \sim K z^\beta e^{-z},\quad z\to\infty,\; \beta\neq-1,\,K>0,$$ ensuring heavy-tailed yet nondegenerate volatility. Under these assumptions, the process possesses a unique strictly stationary solution, with both $\sigma_0$ and $R_0$ being regularly varying of index $-1$ and (for $R_0$) tail-balanced.

Extensions to high-dimensional panels posit for (possibly thresholded) exceedance variables: $$Y_{i,t} = \sigma_{i,t} \cdot \varepsilon_{i,t},\qquad \sigma_{i,t} = l_{0i}' f_{0t},$$ where $\{Y_{i,t}\}$ are “excesses” over an estimated central quantile, $l_{0i}$ and $f_{0t}$ are low-rank loadings and factors ($r\ll N$), and $\varepsilon_{i,t}$ are heavy-tailed idiosyncratic components (Hu et al., 1 Jun 2025).

2. Extremal Dependence and Tail Dependence Coefficient

Unlike classical SV models (e.g., Gaussian-log stochastic volatility with heavy-tailed $Z_t$) that display maximal asymptotic independence ($\eta_\ell \equiv 1/2$ for lag $\ell>0$), the FTVM structure allows the extremal dependence (tail dependence at lag $\ell$)

$$\eta_\ell = \lim_{x\to\infty} \Pr\{R_{t+\ell} > x \mid R_t > x\}$$

to take any value in $[1/2, 1]$. Specifically, $\eta_\ell$ is computed via the minimal value $S_\ell$ of an infinite-dimensional linear program: $$\min_{\{\kappa_i\ge0\}} \sum_{i}\kappa_i \quad\mathrm{s.t.}\; \sum_{i}\alpha_i\kappa_i\geq1,\; \sum_{i}\alpha_{i-\ell}\kappa_i\geq1; \quad \eta_\ell=1/S_\ell.$$ For strictly decreasing $\alpha_i$, $S_\ell=2-\alpha_\ell$ and thus $\eta_\ell=1/(2-\alpha_\ell)$. In the AR(1) case, $\alpha_i=\alpha^i$,

$\eta_\ell = \frac{1}{2-\alpha^\ell}.$

Consequently, FTVM can model strictly decreasing tail dependence coefficients in line with observed clustering of extremes in financial and econometric data (Janssen et al., 2013).

3. Multivariate Regular Variation: Breiman-type Theorems

The probabilistic underpinnings of FTVM extremal dependence rest on two multivariate extensions of classical Breiman's lemma:

• Random matrix times tail vector: If $\mathbf Z\in(0,\infty)^d$ is regularly varying and $\mathbf A$ is a random $d\times d$ matrix satisfying appropriate moment and nondegeneracy conditions, then $\mathbf A \mathbf Z$ is also regularly varying on $(0,\infty)^d$ with the same index [(Janssen et al., 2013), Theorem 2.4].
• Products of i.i.d. regularly varying scalars: For $\{X_i\}$ i.i.d. regularly varying, and exponents $\alpha_i,\beta_i\geq 0$ (summable), the joint exceedance probabilities of products $Y_1=\prod_i X_i^{\alpha_i}$, $Y_2=\prod_i X_i^{\beta_i}$ admit sharp asymptotics via an associated linear program, determining $S=\sum_i\kappa_i$ for the joint tail index [(Janssen et al., 2013), Theorems 4.3–4.4].

These results guarantee the preservation and quantification of heavy-tail phenomena through matrix and product operations as embodied in FTVM factorization and pooling.

4. High-Dimensional Quantile Extension: FTVM–EoT

For panels or high-dimensional arrays $\{X_{i,t}\}$, the FTVM is integrated with the excess-over-threshold (EoT) methodology to form the FTVM–EoT framework (Hu et al., 1 Jun 2025). The procedure is as follows:

1. Central quantile modeling: Fit a high-dimensional quantile model (e.g., QFM or QRIFE) at a central level $\tau^*$, estimating $\hat u_{i,t}$ as threshold.
2. Exceedance formation: Compute $Y_{i,t} = (X_{i,t} - \hat u_{i,t})_+$.
3. Factorization: Model $Y_{i,t} = l_{0i}' f_{0t}\,\varepsilon_{i,t}$, enforcing low-rank volatility structure.
4. Tail fitting: Reference tail quantiles and indices are estimated using order statistics and the Hill estimator.
5. Intermediate and extreme quantile extrapolation: Use estimated factors and GPD-type asymptotics to recover quantiles at levels $k/(NT)$ and $p_{N,T}\ll k/(NT)$.

Algorithmic estimation uses an iterative scheme alternating quantile regressions in factors and loadings. The approach generalizes classical EoT by allowing for dependence and heterogeneity in tail scaling and is supported by formal asymptotic theory under regular variation and identification assumptions.

5. Calibration and Model Selection Procedures

Model selection and validation for FTVM in the high-dimensional context employ two main tools (Hu et al., 1 Jun 2025):

• Degenerate model test: A Kolmogorov–Smirnov statistic on exceedance indicators tests the null hypothesis $H_0$ of no factor-structure ($\sigma_{i,t}\equiv1$). Under $H_0$, the statistic converges to a Brownian bridge supremum; otherwise, factor heterogeneity is declared.
• Information criterion for rank selection: Fit FTVM for ranks $r=0,...,r^*$, compute the quantile loss penalized by an information criterion $P_{N,T}$, and select $\hat r=\arg\min J(r)$. Consistency of $\hat r$ is established as $N,T,k$ diverge with mild penalty growth.

6. Practical Applications and Empirical Evaluation

FTVM and its EoT-integrated extension have been evaluated in simulation experiments:

• In pure FTVM setups, when true data exhibit low-rank tail volatility, the FTVM estimation outperforms degenerate and overfitted models, particularly as sample sizes increase.
• In data generating processes with weak secondary factors, correct model selection by KS test or information criterion is only feasible for large $N,T$.
• Against direct high-dimensional quantile regression (QFM, QRIFE), FTVM–EoT maintains superior mean-squared relative error at both intermediate and extreme quantile levels when tail indices are large (very heavy tails).
• The Hill estimator for the tail index is recommended to be tuned by stability plots due to its bias-variance tradeoff at various threshold choices.

Table: Comparison of Extremal Dependence in SV-type Models

Model	Tail Index $\eta_\ell$	Extremal Dependence
Classical Gaussian-log SV, heavy-tailed	$\eta_\ell=1/2\,\,\forall \ell>0$	Maximal asymptotic independence
GARCH(1,1)	$\eta_\ell=1\,\,\forall \ell$	Full asymptotic dependence
FTVM ($\{\alpha_i\}$ choice)	$\eta_\ell\in[1/2,1]$	Flexible, prescribed by $\{\alpha_i\}$

Simulations support that, for empirically plausible AR(1) log-volatility decay ($\alpha_i = \alpha^i$), the FTVM achieves decreasing $\eta_\ell$, explaining observed tail clustering which is absent in standard models (Janssen et al., 2013, Hu et al., 1 Jun 2025).

7. Applications to Option Pricing and Tail Risk Premia

A related FTVM formulation underlies “Option Pricing, Historical Volatility and Tail Risks” (Vazquez, 2014), where real-world volatility dynamics (estimated via GARCH) are augmented by explicit tail risk premia—convexity ($\lambda_2$), skew ($\lambda_3$), and kurtosis ($\lambda_4$). This leads to a two-stage framework:

1. Estimation of underlying variance dynamics (possibly multi-scale, asymmetric) under $\mathbb{P}$.
2. Augmentation of P&L drift by $(\lambda_2,\lambda_3,\lambda_4)$ according to portfolio sensitivity to large moves, yielding a closed-form fair-pricing PDE for options and a continuous-time SV representation under the pricing measure.

Implied moments and volatility skews are calibrated via weighted integrals of observed option prices. Numerical experiments confirm that introducing nonzero tail premia is essential to replicate observed implied vol surfaces and risk-premia time series. Pure historical volatility underestimates both level and slope without these extensions.

---

References:

• (Janssen et al., 2013) Janssen, A., & Drees, H. (2013). A stochastic volatility model with flexible extremal dependence structure.
• (Hu et al., 1 Jun 2025) Factorized Tail Volatility Model: Augmenting Excess-over-Threshold Method for High-Dimensional Hevay-Tailed Data.
• (Vazquez, 2014) Option Pricing, Historical Volatility and Tail Risks.