Survival Tail Ratio

Statement¶

For a right-tail survival function $\bar F(x)=\mathbb P(X>x)$, the survival tail-ratio diagnostic is

where $\bar F(2x)>0$. With iid copies $X_1$ and $X_2$, the numerator is

Thus $Q_F(x)$ compares a two-coordinate moderate-extreme event with a one-coordinate double-threshold event. It is a tail-geometry diagnostic, not an estimator by itself.

Here $x>0$ is the comparison threshold, $F$ is the CDF, and $\alpha>0$, $x_m>0$ are the Pareto exponent and lower cutoff. $Q_F$ is distinct from the regular-variation ratio $\bar F(tx)/\bar F(x)$.

Calibration cases¶

The diagnostic separates three useful geometries:

The Pareto calculation is exact once $x\ge x_m$. The normal calculation uses Mills’ ratio, $\bar\Phi(x)\sim e^{-x^2/2}/(\sqrt{2\pi}x)$, for the standard Gaussian tail. This also follows from the complementary-error function expansion in NIST DLMF, Section 7.12.

The regular variation connection is simple. If $\bar F\in RV_{-\alpha}$ with $\alpha>0$, then

Therefore

So regularly varying tails fall on the one-big-jump side of this diagnostic. That agrees with subexponentiality, though $Q_F(x)\to0$ is only a diagnostic ratio and not the full convolution-tail definition.

Static calibration¶

For an exact Pareto law with $x_m=1$ and $\alpha=1.5$, $Q_F(x)=(2/x)^{3/2}$ for $x\ge1$. The exponential comparison is exactly one at every positive threshold.

A heavy-tailed model can therefore have $Q_F(x)\ge1$ at finite thresholds; the asymptotic direction and the range inspected both matter. For the normal, substitution of Mills’ ratio gives

Student-t tails with positive degrees of freedom (including the Cauchy case) are regularly varying, so the preceding regular-variation argument gives $Q_F(x)\to0$. These are population calculations, not empirical fits.

Caveats¶

• $Q_F(x)$ is a distribution-level diagnostic. A finite empirical estimate can be dominated by sampling noise because both the numerator and denominator involve rare events.

• The iid interpretation of $\bar F(x)^2$ fails under dependence. Clustered extremes need their own dependence model before this ratio can be read as a joint-event probability.

• The ratio uses right tails. Two-sided returns, losses, and absolute values require an explicit transformation before the diagnostic is applied.

• $Q_F(x)\to0$ is compatible with one-big-jump geometry, but it is not a replacement for the convolution-tail definition of Subexponentiality.

References¶

• Taleb, Statistical Consequences of Fat Tails 2020.

• Feller, An Introduction to Probability Theory and Its Applications, Vol. II (2nd ed., Wiley, 1971).

• Embrechts, Klueppelberg, and Mikosch, Modelling Extremal Events 1997.

Backlinks¶

• Depends on: Regular Variation and Subexponentiality.

• Related catalog: Tail Class Catalog.

• Related diagnostic: Max-to-Sum Ratio.

• Used by: Iso-Density Tail Geometry.

Source and adaptation¶

Adapted from incerto-wiki, content/concepts/methods/survival-tail-ratio.md, revision 9717c9c (2026-09-13 Batch 2 import). Copyright (c) 2023 xshi19. Licensed under MIT. Links, notation, and qualifications were adapted for this site; executable figures and simulations were replaced with static calculations. No upstream execution or formal-proof verification is claimed for this adaptation.