Why Diagnose a TVaR Estimate?
A TVaR estimate can be calculated correctly and still be unreliable. At a high confidence level, only a small part of the loss distribution determines the answer. A fitted tail model may then extrapolate far beyond the observations, making the result sensitive to a few losses, the threshold used for fitting, or the assumed tail shape. Diagnostics do not prove that a model is correct; they show whether the available evidence supports using it.
The useful questions are practical: How many observations can inform the requested tail? Do estimates remain similar when the fitting threshold changes? Does the fitted tail reproduce the observed exceedances? And how much does TVaR move under plausible changes to assumptions? These checks complement estimation methods such as empirical TVaR, parametric fitting, and bootstrap uncertainty assessment.
Start With the Tail-Observation Count
Suppose \(n\) independent losses come from a continuous distribution and \(p\) is the true quantile level. The number \(K\) of observations exceeding \(\operatorname{VaR}_p(X)\) has a binomial distribution with parameters \(n\) and \(1-p\). Thus
The expected count is a first diagnostic, not a guarantee about the sample actually obtained. If the target is the 99.9th percentile in 1,000 observations, the expected number above it is only one. A detailed model fitted to that extreme tail cannot be justified merely by reporting several decimal places. Dependence between observations, changing exposure, or an estimated rather than known quantile can further complicate this simple binomial benchmark.
Worked Example: Is the Target Tail Represented?
A hypothetical insurer has 1,200 independent, identically distributed annual loss observations. Its proposed TVaR level is 99.5%. The quantile is treated as known for this diagnostic calculation. Find the expected number of losses above it, the standard deviation of that count, and the chance that the sample contains no such loss.
Method. Use the binomial count with \(n=1{,}200\) and exceedance probability \(1-p=0.005\).
The expected exceedance count is \(1{,}200(0.005)=6\). Its standard deviation is \[ \sqrt{1{,}200(0.995)(0.005)}=\sqrt{5.97}\approx 2.443. \] The probability of no exceedances is \[ P(K=0)=(1-0.005)^{1{,}200}=0.995^{1{,}200}\approx 0.0024. \]
The sample is likely to contain at least one observation beyond the true quantile, but it is expected to contain only six. Counts will vary materially around six, so a tail estimate based on these observations should be treated as data-limited. The calculation does not establish that six observations are sufficient; it makes the limited support visible. Common mistake avoided: interpreting the expected count as the actual count or as proof of adequate tail information.
Check Stability Across Fitting Thresholds
A high quantile is often estimated from exceedances above a threshold \(u\). For a generalized Pareto distribution (GPD), the conditional survival function of excess \(Y=X-u\), given \(X>u\), is
on the range where the expression is positive. Here \(\beta>0\) is the scale parameter and \(\xi\) is the shape parameter; when \(\xi=0\), the limiting model is exponential with survival \(e^{-y/\beta}\). The threshold exceedance probability is \(s_u=P(X>u)\). For a target \(p\) above the threshold, the fitted quantile is
For \(\xi<1\), the corresponding finite TVaR is \(q_p+\beta_q/(1-\xi)\), where \(\beta_q=\beta+\xi(q_p-u)\) is the GPD scale at \(q_p\). In the exponential case this reduces to \(q_p+\beta\). If \(\xi\ge 1\), the fitted unbounded GPD has infinite mean excess and does not give a finite TVaR.
Under an exact GPD tail, raising the threshold from \(u\) to \(v>u\) preserves the shape parameter and changes the scale to \(\beta_v=\beta+\xi(v-u)\). This threshold-stability property motivates a diagnostic: fit at several plausible thresholds and examine whether the estimated shape, appropriately adjusted scale, and target TVaR form a reasonably stable region. Sampling variation is expected; a systematic drift or a sharp jump deserves investigation.
Worked Example: A Threshold-Stability Check
A hypothetical portfolio has 10,000 independent losses. A GPD tail is fitted at three thresholds. The fitted shape is \(\xi=0\), and the scale is \(\beta=20\) thousand dollars at each threshold. The estimated exceedance probabilities are \(0.1000\) at \(u=100\), \(0.0367879\) at \(u=120\), and \(0.0135335\) at \(u=140\), with thresholds also in thousands of dollars. Estimate 99% TVaR at each threshold and assess stability.
Method. For an exponential excess tail, \(q_{0.99}=u+\beta\ln(s_u/0.01)\) and \(\operatorname{TVaR}_{0.99}=q_{0.99}+\beta\).
At \(u=100\), the ratio \(s_u/(1-p)\) is \(0.1000/0.01=10\). Therefore \(q_{0.99}=100+20\ln(10)=146.0517\), and TVaR is \(146.0517+20=166.0517\) thousand dollars.
At \(u=120\), the ratio is \(3.67879\), whose natural logarithm is approximately \(1.302585\). Thus \(q_{0.99}=120+20(1.302585)=146.0517\), and TVaR is \(166.0517\) thousand dollars.
At \(u=140\), the ratio is \(1.35335\), whose natural logarithm is approximately \(0.302585\). Thus \(q_{0.99}=140+20(0.302585)=146.0517\), and TVaR is again \(166.0517\) thousand dollars. The three estimates are stable under this model, as expected for these threshold probabilities and a constant exponential scale. In practice, fitted values would not usually agree exactly; the counts above the thresholds and the uncertainty in the fitted parameters must also be reviewed. Common mistake avoided: comparing fitted scales alone without checking how the threshold, exceedance probability, and scale combine in the target quantile.
Inspect the Fitted Tail, Not Just Its Summary
A good-looking TVaR value is not a goodness-of-fit test. The fitted model should also be compared with the observed exceedances. A quantile-quantile (Q-Q) plot compares ordered observed excesses with the corresponding fitted GPD quantiles. Points that track a roughly straight diagonal support the fit over the plotted range; curvature, a few extreme departures, or a change in slope can signal tail misfit or influential observations. A probability-probability (P-P) plot compares empirical and fitted exceedance probabilities. It is especially useful for detecting systematic underprediction or overprediction across the excess range.
These plots should be read with the number of exceedances in mind. A smooth-looking plot based on very few points is weak evidence. Also check whether exceedances cluster in time or by exposure, territory, or coverage. The independent, identically distributed count calculation is not an adequate description if the underlying losses are dependent or come from materially different populations. A pooled tail fit may then conceal distinct behavior.
The diagnostics should focus on the portion of the tail that supports the intended estimate. A fit can describe moderate exceedances adequately while extrapolating poorly to a much more extreme quantile. Record both the target tail probability and the tail probability at the fitting threshold; their ratio indicates the extrapolation distance in probability terms. A large ratio is a warning to seek stronger justification, not an automatic rejection rule.
Sensitivity to Tail Shape and Extrapolation
The GPD shape parameter controls how quickly tail risk grows. As a result, modest changes in an estimated shape can materially change an extrapolated TVaR. The following example uses one threshold and one tail family to isolate that sensitivity. The parameter values are hypothetical, not empirical industry estimates.
Worked Example: Shape Sensitivity and a Formula Trap
A hypothetical loss portfolio has 10,000 independent observations. At a threshold of \(u=100\) thousand dollars, 500 observations exceed the threshold, so \(s_u=0.05\). A GPD fit has scale \(\beta=30\) thousand dollars. Estimate 99.5% TVaR under two plausible fitted shapes, \(\xi=0.15\) and \(\xi=0.30\), then assess the sensitivity.
Method. The target exceedance probability is \(1-p=0.005\), so \(s_u/(1-p)=0.05/0.005=10\). For each shape, calculate \(q_p\), update the scale to \(\beta_q\), then calculate \(q_p+\beta_q/(1-\xi)\). Both shapes are less than 1, so the fitted TVaR is finite.
For \(\xi=0.15\), \[ 10^{0.15}\approx1.41254,\qquad q_{0.995}=100+\frac{30}{0.15}(1.41254-1)\approx182.51. \] The scale at this quantile is \(\beta_q=30+0.15(182.51-100)\approx42.38\). Therefore TVaR is \(182.51+42.38/0.85\approx232.36\) thousand dollars.
For \(\xi=0.30\), \[ 10^{0.30}\approx1.99526,\qquad q_{0.995}=100+\frac{30}{0.30}(1.99526-1)\approx199.53. \] Now \(\beta_q=30+0.30(199.53-100)\approx59.86\), so TVaR is \(199.53+59.86/0.70\approx285.04\) thousand dollars. The estimate rises by about \(52.68\) thousand dollars under the higher shape value. The 99.5% tail represents an expected \(10{,}000(0.005)=50\) observations, while the threshold has 500 exceedances; thus the model extrapolates from the threshold to a tail one-tenth as large.
Common mistake and correction. Using \(q_p+\beta/(1-\xi)\) for the second fit gives \(199.53+30/0.70\approx242.39\) thousand dollars. This incorrectly uses the scale at \(u\) to calculate the mean excess above \(q_p\). The correct scale there is \(\beta_q\approx59.86\), giving \(285.04\) thousand dollars. The large difference makes checking the scale update essential.
A Short Justification of the Tail-Mean Formula
The formula for TVaR above a GPD quantile follows from the mean excess remaining at that quantile. Conditional on exceeding \(q_p\), the excess \(X-q_p\) has a GPD tail with scale \(\beta_q\) and shape \(\xi\), by the threshold stability property. For a nonnegative excess, its expectation is the integral of its survival function:
For \(\xi>0\), the integral converges only when \(1/\xi>1\), or \(\xi<1\); for \(\xi=0\), integrating \(e^{-y/\beta_q}\) gives \(\beta_q\). For \(\xi<0\), the excess has a finite upper endpoint and integration over its support gives the same expression when \(\xi<1\). Adding the starting loss \(q_p\) gives \(\operatorname{TVaR}_p=q_p+\beta_q/(1-\xi)\). This also explains why a fitted shape at or above 1 is a warning that the GPD model implies infinite TVaR, not merely a large finite estimate.
Turn Diagnostics Into a Judgment
No single diagnostic establishes reliability. A defensible assessment combines data support, fit, and sensitivity. In particular, report the number of threshold exceedances and the expected number beyond the target quantile; show how the TVaR estimate changes across reasonable thresholds; inspect plots and any systematic departures; and test sensitivity to plausible parameter changes. Where available, uncertainty intervals provide additional context, but they do not repair a biased or poorly supported tail model.
- Few target-tail observations: acknowledge that the estimate depends heavily on the model rather than direct observations.
- Strong threshold drift: investigate whether the threshold is too low, the tail is not well described by the fitted model, or the data combine different loss processes.
- Shape sensitivity: present the range of results and explain which assumptions drive it rather than presenting one estimate as definitive.
- Poor plot fit or clustered losses: examine model form and dependence before treating conventional independent-sample diagnostics as decisive.
- Nonfinite fitted mean: if a GPD shape estimate is at least 1, the fitted model does not support a finite TVaR.
The purpose of these checks is not to select whichever fit gives a preferred capital number. It is to make the strength and limits of the estimate transparent. This diagnostic discipline is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.
Check Your Understanding
For each problem, identify what the calculation says about support for the TVaR estimate, not just its numerical result.
- A hypothetical sample contains 2,400 independent losses, and the target level is 99%. Find the expected number of observations beyond the true quantile and the standard deviation of that count. (Answer: Expected count \(=24\); standard deviation \(\approx4.874\).)
- A fitted threshold has estimated exceedance probability \(0.08\), and the target TVaR level is 99%. What is the ratio of threshold exceedance probability to target exceedance probability? Interpret it as an extrapolation factor in probability terms. (Answer: \(0.08/0.01=8\); the target tail is one-eighth the threshold exceedance probability.)
- An exponential excess model has threshold \(u=60\) thousand dollars, scale \(\beta=15\) thousand dollars, and threshold exceedance probability \(0.04\). Estimate 99% TVaR. (Answer: \(q_{0.99}=60+15\ln(4)\approx80.79\); \(\operatorname{TVaR}_{0.99}\approx95.79\) thousand dollars.)
- A GPD fit has \(u=200\) thousand dollars, \(\beta=25\) thousand dollars, \(\xi=0.20\), and threshold exceedance probability \(0.10\). Estimate 99% TVaR, showing the scale update at the target quantile. (Answer: \(q_{0.99}\approx273.11\), \(\beta_q\approx39.62\), and \(\operatorname{TVaR}_{0.99}\approx322.64\) thousand dollars.)
- A hypothetical data set has 20,000 independent losses. A GPD tail is fitted above a threshold exceeded by 2% of losses, while the target level is 99.9%. Find the expected number beyond the target quantile and the probability extrapolation ratio. What diagnostic concern should be stated? (Answer: Expected target-tail count \(=20\); ratio \(0.02/0.001=20\). The estimate extrapolates from the threshold tail to a tail with one-twentieth its probability, with only about 20 expected observations beyond the target quantile.)