Estimating Dependence with Sparse Extreme Observations

Why Sparse Extremes Require Caution

Tail-dependence estimates are built from observations in the far ends of both marginal distributions. Those observations are precisely the ones a finite data set contains least often. A reported estimate can therefore look precise while depending on only a handful of joint exceedances. The estimate should be accompanied by an uncertainty assessment and a check of how it changes when the threshold changes.

The focus here is empirical upper-tail dependence. The same reasoning applies to lower-tail dependence after reversing the direction of the inequalities. Assume paired observations $(X_i,Y_i)$, $i=1,\ldots,n$, are independent and identically distributed, with continuous marginal distributions. Transforming each margin to uniform ranks gives pseudo-observations $U_i,V_i$ on the unit interval. For example, one common convention is $U_i=R_{X_i}/(n+1)$, where $R_{X_i}$ is the rank of $X_i$; use the same convention for $V_i$.

Definition: For a copula $C$ with continuous uniform margins, the upper-tail dependence coefficient is $$ \lambda_U=\lim_{q\uparrow 1}P(U>q\mid V>q) =\lim_{q\uparrow 1}\frac{1-2q+C(q,q)}{1-q}, $$ when the limit exists. At a fixed threshold $q<1$, the conditional exceedance probability $P(U>q\mid V>q)$ is a finite-threshold quantity; it is not automatically equal to the limiting coefficient $\lambda_U$. An empirical estimate at a high threshold describes the observed tail at that threshold and is often used as an approximation to the limit.

For an empirical estimate, let $m$ be the number of observations exceeding $q$ in the conditioning margin $V$, and let $x$ be the number that also exceed $q$ in $U$. Then $\hat p_q=x/m$ estimates the finite-threshold conditional probability. With independent paired observations and a fixed threshold, conditional on which observations exceed $q$ in $V$, the $m$ indicators of whether $U>q$ are Bernoulli trials with success probability $p_q=P(U>q\mid V>q)$. Thus $x$ has a binomial distribution with size $m$ and probability $p_q$. This binomial model is useful for uncertainty at a fixed threshold, but it does not account for estimating the threshold from the data or for the difference between $p_q$ and the limiting coefficient.

A Confidence Interval for a Sparse Count

A normal approximation interval can behave poorly when the count is small, especially when the estimated probability is near zero or one. The Wilson score interval is a practical alternative. It is obtained by inverting the binomial score test and generally stays within the probability range from zero to one.

Result: For $x$ joint exceedances among $m$ conditioning exceedances, write $\hat p=x/m$. A two-sided Wilson interval with standard-normal critical value $z$ has center $c$ and half-width $h$: $$ c=\frac{\hat p+z^2/(2m)}{1+z^2/m}, \qquad h=\frac{z}{1+z^2/m} \sqrt{\frac{\hat p(1-\hat p)}{m}+\frac{z^2}{4m^2}}. $$ The interval is $[c-h,c+h]$. For a 95% interval, use $z=1.96$. It estimates uncertainty in the finite-threshold conditional probability under the binomial assumptions; it is not, by itself, a confidence interval for the asymptotic tail-dependence coefficient.

To see why the expression has this form, suppose a candidate probability $p$ is not rejected by the binomial score test. The score condition is $$ \frac{(\hat p-p)^2}{p(1-p)/m}\le z^2. $$ Multiplying by the positive denominator and collecting terms gives a quadratic inequality in $p$. Its two roots are the Wilson endpoints. Completing the square expresses those roots as $c-h$ and $c+h$ above. The adjustment terms involving $z^2$ are important when $m$ is small; omitting them produces the ordinary Wald interval, which can be unreliable in precisely this sparse setting.

Worked Example: A 95% Interval From Ten Tail Observations

In a hypothetical portfolio study, 100 paired annual loss observations are converted to uniform pseudo-observations. At threshold $q=0.90$, exactly 10 observations exceed the threshold in the conditioning margin, and 3 of those also exceed it in the other margin. Estimate the finite-threshold conditional probability and calculate a 95% Wilson interval.

Method. Use the joint-exceedance count as the binomial success count: $x=3$, $m=10$. The point estimate is $$ \hat p=\frac{x}{m}=\frac{3}{10}=0.3000. $$

For the interval, $z=1.96$, so $z^2=3.8416$. The common denominator is $$ 1+\frac{z^2}{m}=1+\frac{3.8416}{10}=1.38416. $$ The center is $$ c=\frac{0.3000+3.8416/(2\cdot10)}{1.38416} =\frac{0.49208}{1.38416}\approx0.3555. $$ The square-root term is $$ \sqrt{\frac{0.3000(0.7000)}{10}+\frac{3.8416}{4(10)^2}} =\sqrt{0.021000+0.009604} =\sqrt{0.030604}\approx0.1749. $$ Thus $$ h=\frac{1.96(0.1749)}{1.38416}\approx0.2477. $$ The interval is $[0.3555-0.2477,\;0.3555+0.2477]$, or approximately $[0.1078,\;0.6032]$. Rounded to four decimals, the estimate is 0.3000 and the 95% interval is $[0.1078,\;0.6032]$. The endpoints bracket the point estimate, as required.

Interpretation and trap. The estimate is based on only 10 conditioning exceedances, and the interval is correspondingly broad. Calling 0.3000 a precise estimate of the limiting coefficient $\lambda_U$ would confuse a finite-threshold rate with a tail limit. Also, using all 100 observations as the binomial denominator would incorrectly give $3/100$: the conditional probability is based on observations that exceed in the conditioning margin.

Paired Bootstrap for Tail Estimates

A paired nonparametric bootstrap assesses how the estimate varies when the observed paired sample is perturbed. Each bootstrap sample draws $n$ row indices with replacement from the original data. Both margins must be resampled together: sampling the $X$ and $Y$ observations separately breaks their observed association and can make the resampling distribution irrelevant to the dependence estimate.

Resample pairs.
Draw $n$ paired rows with replacement, not separate samples from each margin.

Rebuild the rank-scale data.
Within the resample, recompute the marginal ranks and pseudo-observations if the original estimator used empirical margins.

Recalculate the estimator.
At the chosen threshold, count conditioning exceedances and joint exceedances, then calculate their ratio. Use the same threshold rule in every replicate.

Summarize the replicates.
Repeat many times. Percentiles of the bootstrap estimates can give a percentile interval; report the replicate count and the threshold rule used.

The paired bootstrap reflects sampling variability in the observed data and, when ranks and empirical thresholds are recomputed, variability from those steps as well. It does not remove bias from using a threshold that is too low to represent the limiting tail, nor does it account automatically for choosing a copula model. If observations are serially dependent, ordinary row-by-row resampling is generally inappropriate; a block bootstrap that resamples consecutive stretches may be needed to preserve temporal dependence.

Worked Example: Reading a Paired Bootstrap Interval

A hypothetical analyst has 80 paired observations and estimates upper-tail dependence at the empirical 90th-percentile threshold. The estimate is $2/8=0.2500$. The analyst performs 1,000 paired bootstrap resamples, recomputes ranks and the threshold in each resample, and sorts the 1,000 resulting estimates from smallest to largest. The 25th and 975th sorted values are 0.0800 and 0.4600. Use these values to form and interpret a 95% percentile interval.

Method. For 1,000 sorted replicates, the lower and upper 2.5% tails correspond approximately to the 25th and 975th ordered estimates under the stated indexing convention. The percentile interval is therefore $$ [0.0800,\;0.4600]. $$ The original estimate, 0.2500, lies inside this interval. Its width is $0.4600-0.0800=0.3800$, indicating substantial resampling variability relative to the point estimate.

Interpretation and trap. The interval summarizes variation under the specified paired-resampling procedure and threshold rule. It is not evidence that the true tail coefficient must lie between these endpoints, particularly if the finite threshold is a poor approximation to the limit. Resampling the two columns independently would be a serious mistake: it would discard the paired association whose tail behavior is being estimated.

Threshold Sensitivity and What It Can Reveal

A lower threshold provides more observations and often a more stable estimate, but it may describe ordinary high values rather than the limiting tail. A higher threshold is more relevant to extreme behavior, but leaves fewer observations and wider uncertainty. Comparing several thresholds makes this trade-off visible. It is a sensitivity check, not a procedure that proves which threshold is correct.

Worked Example: Comparing Three Upper-Tail Thresholds

For a hypothetical sample of 200 paired observations, an analyst calculates the empirical conditional exceedance rate at three thresholds. The observed counts are: at $q=0.90$, 20 conditioning exceedances and 7 joint exceedances; at $q=0.95$, 10 and 4; and at $q=0.975$, 5 and 2. Compare the estimates and assess what the counts imply.

Step 1: Calculate each rate. The three estimates are $$ \hat p_{0.90}=\frac{7}{20}=0.3500,\qquad \hat p_{0.95}=\frac{4}{10}=0.4000,\qquad \hat p_{0.975}=\frac{2}{5}=0.4000. $$

Step 2: Check the nested counts. The more extreme exceedance sets are subsets of the less extreme sets: 5 is no greater than 10, which is no greater than 20; likewise 2 is no greater than 4, which is no greater than 7. The counts are internally consistent with nested thresholds.

Step 3: Interpret the pattern. The point estimates are similar, ranging from 0.3500 to 0.4000, but the most extreme estimate uses only five conditioning observations. Similar point estimates do not imply similar precision. In particular, one additional joint exceedance at the highest threshold would change its estimate from $2/5=0.4000$ to $3/5=0.6000$.

The practical conclusion is not that the coefficient is known to be 0.4000. The threshold comparison provides some evidence that the observed rates are stable across these choices, while the very small extreme count leaves substantial uncertainty. Treating the three estimates as independent evidence or selecting the threshold solely because its point estimate looks appealing would be misleading.

Common Mistakes and Practical Reporting

Reporting a tail estimate without its count. State both the estimate and the number of conditioning and joint exceedances; a ratio based on 2 of 5 observations deserves different caution from one based on 200 of 500.
Calling a finite-threshold rate the tail limit. Label the observed estimate by its threshold and explain that it approximates $\lambda_U$ only if the finite-threshold behavior is informative about the limit.
Resampling margins separately. This destroys the pairing. Resample complete rows and, where appropriate, recompute ranks and thresholds in each replicate.
Ignoring dependence across time. The binomial interval and ordinary paired bootstrap rely on independent observations. For serially dependent data, use an approach that preserves relevant time dependence and state its assumptions.
Using a narrow interval as a model guarantee. A confidence interval or bootstrap interval quantifies uncertainty under its sampling procedure. It does not resolve threshold choice, extrapolation to the limit, or copula misspecification.

Key takeaway: Sparse joint extremes make tail-dependence estimates unstable. Report the threshold and exceedance counts, quantify sampling uncertainty with a suitable interval or paired bootstrap, and check whether conclusions are sensitive to the threshold. Keep finite-threshold evidence separate from claims about the asymptotic coefficient.

Check Your Understanding

For each problem, state the estimate and interpret it in light of the specified information.

At a fixed upper threshold, 16 observations exceed in the conditioning margin and 5 exceed jointly. Calculate the empirical conditional exceedance rate. (Answer: $5/16=0.3125$.)
A study has 12 conditioning exceedances, of which 1 is joint. Calculate the finite-threshold estimate and state why it should not automatically be reported as the limiting tail coefficient. (Answer: $1/12=0.0833$; it is a finite-threshold estimate based on only 12 conditioning observations, not automatically the limit $\lambda_U$.)
A paired bootstrap produces 1,000 sorted estimates. Under the stated percentile convention, the 25th value is 0.12 and the 975th is 0.51; the original estimate is 0.30. Give the 95% percentile interval and say whether the original estimate lies inside it. (Answer: $[0.12,0.51]$; yes, 0.30 lies inside.)
At two nested thresholds, the counts are 18 conditioning and 6 joint exceedances at the less extreme threshold, and 8 conditioning and 3 joint exceedances at the more extreme threshold. Calculate both rates and explain the precision trade-off. (Answer: $6/18=0.3333$ and $3/8=0.3750$; the rates are similar, but the more extreme estimate uses fewer observations and is less precise.)
A data set consists of 400 consecutive daily paired loss observations with evidence of serial dependence. An analyst proposes an ordinary paired bootstrap that resamples individual days. Identify a concern and a more suitable resampling approach. (Answer: resampling individual days can break serial dependence; consider a paired block bootstrap that resamples consecutive stretches.)

What You'll Learn