Quantile Forecast Vincentization

• Quantile-based forecast Vincentization is a method that averages model quantiles at each probability level to construct a coherent predictive distribution.
• It employs a two-step estimation procedure that first minimizes quantile loss and then optimizes a joint VaR–ES loss using Beta-weighted aggregation.
• The approach effectively mitigates model uncertainty and adapts to heavy-tailed, non-Gaussian data, enhancing risk estimation in complex, high-dimensional settings.

Quantile-based forecast Vincentization is a methodology for probabilistic forecast combination that constructs aggregate quantile forecasts by optimally averaging component model quantiles at each probability level, then (optionally) pools these across quantile levels to estimate risk functionals such as Expected Shortfall (ES). Originally developed for financial tail-risk applications, quantile-based Vincentization systematically addresses model uncertainty and enables robust, coherent estimation of quantile-dependent targets including Value-at-Risk (VaR) and ES. The foundational procedure leverages a two-step estimation: first, minimizing quantile (check) loss for combined quantile forecasts; second, minimizing a strictly consistent joint VaR–ES loss within the Fissler–Ziegel framework for coherent ES aggregation. Extensions to high-dimensional, non-Gaussian settings with quantile factor models further generalize the approach to broader forecasting contexts (Storti et al., 2021, Park et al., 2024).

1. Foundational Definition and Principle

Vincentization, in probabilistic forecasting, refers to averaging a set of predictive distribution functions or—in the quantile domain—their inverse-quantile functions. Quantile-based forecast Vincentization applies this concept at each probability level $\tau \in (0,1)$. Given $n_\mathrm{mod}$ models and the one-step-ahead $\tau$-quantile forecast $q_{t,i}(\tau)$ from model $i$, the Vincentized quantile forecast is defined as

$$Q_t^{(C)}(\tau) = \sum_{i=1}^{n_\mathrm{mod}} w_i(\tau)\,q_{t,i}(\tau)$$

with weights $w_i(\tau) \geq 0$ and $\sum_{i=1}^{n_\mathrm{mod}} w_i(\tau) = 1$. The weights may vary with $\tau$, allowing flexible pooling of model-specific strengths at different distribution regions or data tails (Storti et al., 2021). This form ensures that the combined forecast remains a valid quantile function at each $\tau$.

In high-dimensional forecasting, structurally analogous schemes are realized by combining quantile forecast predictions from quantile factor models, with aggregation weights possibly assigned based on data-driven state probabilities or transition dynamics (Park et al., 2024).

2. Mathematical Formulation and Loss Functions

At a given quantile level $\tau$, the combined quantile is often constructed with a possible intercept: $$Q_t^{(C)}(\tau; c(\tau)) = c_0(\tau) + \sum_{i=1}^{n_\mathrm{mod}} c_i(\tau) q_{t,i}(\tau)$$ subject to non-negativity and sum-to-one constraints on $c_i(\tau)$. Coefficients $c(\tau)$ are estimated by minimizing the empirical average quantile loss (check or pinball loss) over a rolling window of $N$ observations: $$\overline{L}_Q(\tau; c(\tau)) = \frac{1}{N} \sum_{s=1}^N L_Q(\tau; y_{t-s}, Q_{t-s}^{(C)}(\tau; c(\tau)) )$$ where

$L_Q(\tau; y, q) = (1_{\{y < q\}} - \tau) (q - y).$

This loss formulation directly targets correct quantile coverage and robustness to outliers. At each $t$ and $\tau$, the coefficients are obtained by

$$\hat c(\tau) = \arg\min_{c(\tau)}\; \frac{1}{N}\sum_{s=1}^N \left(1_{\{y_{t-s} < c_0(\tau)+\sum_i c_i(\tau) q_{t-s,i}(\tau)\}} - \tau \right)\, \left( c_0(\tau)+\sum_i c_i(\tau) q_{t-s,i}(\tau) - y_{t-s} \right).$$

(Storti et al., 2021).

3. Two-Step Estimation Procedures

Step 1: Quantile Vincentization

• Grid selection: Choose $M$ quantile levels $0<\alpha_1<\cdots<\alpha_M=\alpha$ (e.g., $\alpha = 0.025$).
• Quantile combination: For each $\alpha_j$, combine model quantile forecasts over the rolling window by minimizing the quantile loss, producing $Q_t^{(C)}(\alpha_j)$, $j=1,\ldots,M$.

Step 2: Weighted Quantile Expectation for ES

The ES at level $\alpha$ admits representation

$$ES_t(\alpha) = \frac{1}{\alpha} \int_0^\alpha Q_t(p)\,dp,$$

which, on a discrete grid, is approximated as

$$ES_t \approx w_0 + \sum_{j=1}^M w_j Q_t^{(C)}(\alpha_j),$$

with $\sum_j w_j = 1$, $w_j \geq 0$, and $w_0$ for left-truncation bias. Weights $w_j$ are derived from a Beta-density, $w_j = B(j/M;a,b)$, estimated by minimizing a strictly consistent joint VaR–ES loss, such as the Asymmetric-Laplace (AL) log-score: $$S_t(y_t, Q_t, ES_t) = -\log((\alpha-1)/ES_t) - \frac{(y_t-Q_t)(\alpha - 1_{\{y_t<Q_t\}})}{\alpha\, ES_t},$$ where $Q_t=Q_t^{(C)}(\alpha)$ is the combined quantile at the target level (Storti et al., 2021).

4. Strictly Consistent Joint VaR–ES Loss Functions

The Fissler–Ziegel class comprises all strictly consistent scoring rules for simultaneous VaR and ES estimation. Any such function can be written (up to functions of $y_t$) as

$$S_t(y, Q, ES) = (1_{\{y<Q\}} - \alpha) G_1(Q) - 1_{\{y<Q\}} G_1(y) + G_2(ES) \Bigl(ES - Q + \frac{1_{\{y<Q\}}}{\alpha}(Q-y)\Bigr) - H(ES) + a(y),$$

for suitable monotone $G_1, G_2$, and $H' = G_2$. The AL log-score used in (Storti et al., 2021) is recovered by $G_1 \equiv 0$, $G_2(x) = -1/x$, $H(x) = -\ln(-x)$. Minimizing any $S_t$ jointly over the rolling window ensures the aggregation weights are valid for tail-risk backtesting.

5. Implementation and Model Universes

Vincentization requires selecting a grid $\{\alpha_j\}$ and model universe. In financial tail risk, models may include parametric GARCH-type VaR predictors and semi-parametric tail-quantile models. For example, (Storti et al., 2021) uses GJR-GARCH-t, EGARCH-t, POT-GJR, POT-EGARCH, GJR-GARCH-HS, EGARCH-HS, CAViaR-AS, and CARE-AS. The practical procedure iterates rolling-window quantile regression and joint VaR–ES loss minimization with explicit pseudocode as follows:

Input: returns {y_1,…,y_T}, in-sample size N, quantile grid {α_j}_{j=1}^M, n_mod models producing q_{t,i}(α_j) for each t,i,j. For t = N+1 … T: For each grid level j=1…M: — Build past window y_{t-N},…,y_{t-1} and model quantiles q_{s,i}(α_j). — Solve c^(j) = argmin_c (1/N)∑_{s=1}^N L_Q(α_j; y_{t-s}, c_0 + ∑ c_i q_{t-s,i}(α_j)). — Compute combined quantile Q_t^{(C)}(α_j) = c_0 + ∑ c_i q_{t,i}(α_j). End j — Let Q_t ≡ Q_t^{(C)}(α_M)=VaR(α). — Solve θ = argmin_{w_0,a,b} (1/N)∑_{s=1}^N S(y_{t-s}, Q_{t-s}, ES_{t-s}(w_0,a,b)) where ES_{t-s}(w_0,a,b) = w_0 + ∑_{j=1}^M B(j/M;a,b) Q_{t-s}^{(C)}(α_j). — Compute one-step-ahead ES: ES_t = w_0 + ∑ B(j/M;a,b) Q_t^{(C)}(α_j). End t

(Storti et al., 2021).

6. Extensions to High-Dimensional Non-Gaussian Data

Quantile-based forecast Vincentization generalizes to settings with high-dimensional predictors and non-Gaussian distributions by leveraging quantile factor models:

• At each quantile level $\tau$, fit a quantile factor model $$Q_{x_{it}}(\tau \mid f_{t,\tau}) = \lambda_{i,\tau}' f_{t,\tau}$$ for predictors $X_t$, minimizing quantile check loss by alternating minimization.
• Fit quantile regression of the target on the resulting $\tau$-factor, yielding $\hat q_{T+h}(\tau)$ as the $\tau$-quantile forecast.
• Combine quantile forecasts $\{\hat q_{T+h}(\tau_\ell)\}_{\ell=1}^m$ using weights $w_{\ell,T}$ derived from a Markov chain over past quantile-based states of the target: $$\hat y_{T+h} = \sum_{\ell=1}^m w_{\ell,T} \hat q_{T+h}(\tau_\ell),$$ where $w_{\ell,T} = P_{k\ell}$ and $S_T = k$ is the current estimated quantile interval of the target (Park et al., 2024).

Key steps include quantile grid and factor number selection, estimation of Markov transition matrices for state-dependent weights, and combination of quantile-specific forecasts. This approach ensures adaptability to heavy-tailed and skewed data seen in environmental and econometric time series.

7. Statistical and Practical Advantages

Pooling multiple model quantiles at each probability level reduces specification risk; if one model is locally misspecified, other models compensate. Optimization of the check-loss tunes weights for tail coverage accuracy. The subsequent Beta-weighted ES estimation builds in optimal quantile-averaging, produces a strictly consistent (coherent) ES estimator, and supports rigorous backtesting (Storti et al., 2021). Empirically, the two-step "Forecast Combination + Weighted Quantile" (FC-WQ) procedure yields more stable and accurate VaR and ES forecasts, particularly in the presence of structural breaks or extreme events.

For high-dimensional non-Gaussian series, Vincentized quantile forecasting demonstrates robust adaptation to skewed, heavy-tailed distributions, outperforming mean-based factor models and standard AR/ARIMA in simulation and real-world pollution data (Park et al., 2024). Limitations include computational demands, the requirement for judicious quantile grid and factor number selection, and potential ad-hoc nature of state-probability weighting.

---

References

• [Modelling uncertainty in financial tail risk: a forecast combination and weighted quantile approach, (Storti et al., 2021)]
• [Combined Quantile Forecasting for High-Dimensional Non-Gaussian Data, (Park et al., 2024)]