Extrapolating Extreme Risk Measures via Asymptotic Tail Proportionality

30 concepts18 papers4 proposals

When managing catastrophic risks—such as financial market crashes, massive insurance claims, or once-in-a-century environmental disasters—we are forced to estimate risk thresholds far beyond the range of historically observed data. If we want to calculate a risk measure at an extreme confidence level (say, the $99.99\%$ Value-at-Risk from a sample of only a few thousand days), an empirical approach fails completely: there are simply no data points available in that region of the distribution tail.

To solve this problem safely, modern risk statistics relies on intermediate-to-extreme tail extrapolation. The core intuition behind this framework is that while the extreme tail is a data desert, the intermediate tail—a region far enough out to reflect the distribution's asymptotic behavior, yet close enough to the center to contain a sparse but usable number of observations—is rich with structural information. By establishing precise mathematical relationships within this intermediate zone and leveraging Extreme Value Theory (EVT), statisticians can reliably project risk metrics into the unobserved extreme tail.

The Bridge: Asymptotic Tail Proportionality

The fundamental reason we can extrapolate into the dark is that under heavy-tailed distributions (often modeled via Pareto-type behavior or regular variation), different risk measures become mathematically locked to one another in the far tail. They exhibit asymptotic tail proportionality.

Consider standard quantiles ($L^1$ optimization) and expectiles ($L^2$ optimization). While they capture different aspects of risk—quantiles look only at the probability of exceeding a threshold, whereas expectiles are sensitive to the actual magnitude of the losses beyond that threshold—their relationship stabilizes as we move deeper into the tail. As shown by Abdelaati Daouia et al. in Extreme M-quantiles as risk measures: From $L^{1}$ to $L^{p}$ optimization, risk measures derived from different loss functions are asymptotically proportional in the distribution tails.

This proportionality means that a highly complex or otherwise unestimable risk metric can be expressed as a simple scaling of a standard extreme quantile. For instance, Gilles Stupfler et al. leveraged first-order regular variation theory in Asymptotic Properties of Generalized Shortfall Risk Measures for Heavy-tailed Risks to prove that sophisticated, rank-dependent utility risk measures maintain a stable proportional relationship with extreme quantiles. Similarly, Simone A. Padoan and Gilles Stupfler utilized this exact principle in Joint inference on extreme expectiles for multivariate heavy-tailed distributions to construct Quantile-Based (QB) estimators that facilitate joint inference across multiple extreme risk dimensions.

  Data-Rich Region          Intermediate Tail             Extreme Tail
(Empirical Estimation)     (Sparse but Present)       (No Historical Data)
       [======] ----------> [Intermediate Metric] ----> [Extreme Metric]
                                    |                         ^
                                    | (Tail Index Scaling)    |
                                    +-------------------------+

The Mechanism: Weissman-Type Extrapolation

How do we practically operationalize this intuition? The standard workhorse is the Weissman-type estimator, which acts as a scaling elevator from the intermediate tail to the extreme tail. The procedure follows three distinct steps:

  1. Intermediate Estimation: Calculate the target risk measure (or a baseline quantile) at an intermediate level $\alpha_k = 1 - k/n$, where the number of exceedances $k$ grows with the sample size $n$ but at a slower rate ($k \to \infty, k/n \to 0$). At this level, data points are sparse but sufficient to compute a stable empirical or semi-parametric estimate.
  2. Tail Index Estimation: Estimate the tail index $\gamma$ (typically using a Hill-type estimator), which characterizes the power-law decay and heaviness of the distribution's tail. This parameter dictates the "speed" at which risks grow as we move further out.
  3. Power-Law Projection: Extrapolate to the extreme level $\alpha_n$ (where $\alpha_n \to 1$ and $n(1-\alpha_n) \to c \ge 0$) by scaling the intermediate estimate by the ratio of the tail probabilities raised to the power of the tail index:

$$\hat{q}(\alpha_n) = \hat{q}(\alpha_k) \cdot \left( \frac{1 - \alpha_k}{1 - \alpha_n} \right)^{\hat{\gamma}}$$

This extrapolation framework underpins much of modern extreme value analysis. For example, Stéphane Girard and Gilles Stupfler extended this concept to multi-dimensional spaces in Extreme geometric quantiles in a multivariate regular variation framework, laying the groundwork for multivariate Weissman-type estimators. In time-series contexts where observations are serially dependent, A. C. Davison et al. demonstrated in Tail Risk Inference via Expectiles in Heavy-Tailed Time Series that combining intermediate expectile estimators with Weissman-type constructions provides a principled, low-variance alternative to naive empirical methods.

Adapting to Bounded and Conditional Environments

While the power-law scaling of heavy tails represents the classic paradigm, real-world data frequently introduces structural complexities like boundaries, dependencies, and streaming inputs. The intermediate-to-extreme extrapolation framework has evolved to handle these scenarios through several key innovations:

Short-Tailed and Bounded Distributions

Not all financial or environmental assets exhibit infinite heavy tails; some possess a theoretical maximum or minimum value. In Extreme expectile estimation for short-tailed data, Abdelaati Daouia et al. developed an asymptotic expansion that maps the relationship between extreme expectiles, standard quantiles, and the finite endpoint of a distribution. This allows risk managers to safely extrapolate within bounded environments without misclassifying them as heavy-tailed.

Regression and Auxiliary Data

Often, extreme risk depends heavily on covariates (e.g., predicting extreme river levels based on upstream rainfall). Early foundations established by the authors of Extremal quantile regression combined linear quantile regression with extreme value constraints to ensure tail equivalence across different regressor values.

More recently, non-parametric methods have removed rigid linear assumptions. In Nonparametric extreme conditional expectile estimation, Stéphane Girard et al. introduced a kernel-smoothing approach to estimate intermediate conditional expectiles before applying a Weissman-type projection. This was further expanded to high-dimensional and continuous curve predictors by Stéphane Girard et al. in Functional estimation of extreme conditional expectiles.

Online Streaming Data

In high-frequency environments, recalculating extreme quantiles over massive historical datasets is computationally prohibitive. To address this, Jian Wang et al. proposed an online algorithm in Online prediction of extreme conditional quantiles via B-spline interpolation. Their approach couples Generalized Pareto Distribution (GPD) tail extrapolation with B-spline interpolation, allowing the model to update its extreme tail estimates continuously as data streams in.

By treating the tail not as an arbitrary sequence of points but as a highly structured geometric profile governed by regular variation, intermediate-to-extreme extrapolation transforms the impossible task of counting unobserved events into a tractable problem of scaling and geometry.

Key Foundational Work

Extremal quantile regression (2005) — contributed 2 concept elements to this direction.
Extreme geometric quantiles in a multivariate regular variation framework (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_extreme_value_extrapolation_74825c95,
  title = {extreme value extrapolation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extreme-value-extrapolation/rdr_74825c95528e4f83bc9bec965e96d003},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_extreme_value_extrapolation_74825c95_v1,
  title = {extreme value extrapolation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extreme-value-extrapolation/rdr_74825c95528e4f83bc9bec965e96d003/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake