MR 3 – Parametric Approaches II (Extreme Values)

A stock market crash, the collapse of a large bank, a sudden financial crisis or a natural catastrophe each happens rarely, yet any one of them can cost more than years of ordinary trading losses. Risk managers call these low-probability, high-impact events, and much of what a risk function exists to do is put a number on them before they arrive. The awkward part is that the quantity we most want to estimate lives exactly where our data are thinnest.

Think about what a five-year daily return history actually contains. It holds roughly 1,250 observations, but only a dozen or so of them sit in the extreme left tail, and the very worst day appears once. An estimate built on so few points is bound to be uncertain, and that uncertainty grows sharply when the question reaches past the data altogether, for instance when a regulator asks how large a loss could be in a fall worse than anything recorded so far. No amount of clever arithmetic can create extreme observations that never happened.

Why fitting one distribution to all the data fails in the tail

The usual workaround is to pick a familiar distribution, often the normal, and fit it to the whole sample. The fitting procedure then does what it is designed to do: it matches the bulk of the data, because the central observations vastly outnumber the extreme ones and dominate any likelihood or least-squares criterion. A model tuned to the middle of the distribution can describe everyday fluctuations well and still be badly wrong about the size of a one-in-a-thousand-day loss, which is the very number the risk manager needs.

Other fields met this problem long before finance did. Hydrologists designing dikes and sea walls must decide how high a barrier has to be so that the chance of a flood stays acceptably small, and they usually work with even shorter records than a trading desk and with far more at stake. Their target water levels typically lie well outside anything observed, so they were forced to develop a statistics of extremes rather than of averages. Insurers face the same task when pricing catastrophe cover.

Extremes follow their own limit theorems

The result is extreme value theory (EVT), a branch of applied statistics built specifically for rare events. Ordinary statistics leans on the central limit theorem, which tells us how sums and averages behave as samples grow. That theorem says nothing useful about the largest observation in a sample. Extremes are governed instead by extreme value theorems, which identify the limiting distributions that maxima and tail exceedances must follow. EVT uses these results in two ways: it tells us which distributions are legitimate to fit to extreme data (and which are not), and it guides how the parameters of those distributions should be estimated. The resulting distributions look and behave differently from the normal family, and their parameters are harder to estimate.

The lesson concentrates on VaR, with some attention to expected shortfall (ES), but the same machinery extends to other coherent risk measures. Univariate EVT comes in two main branches. The first, generalized extreme value theory, models the maximum of a large sample. The second, the peaks-over-threshold approach, models the losses that exceed a high threshold. After both branches come refinements for conditional and dependent data, a look at the multivariate case, and an honest account of the limits.

Let X be a random loss, assumed for now to be independent and identically distributed (iid) with some unknown distribution function F(x) = Prob(X ≤ x). Because F is unknown, we cannot simply read an extreme quantile off it. EVT sidesteps the problem by changing the object of study. Draw a sample of size n from F and record its largest value, M_n. When n is large, M_n is itself an extreme value, and the question becomes: what distribution does M_n follow?

The Fisher-Tippett theorem (1928) supplies the answer under fairly general conditions. As n grows, the distribution of M_n, suitably located and scaled, converges to a single three-parameter family, the generalized extreme value (GEV) distribution:

The striking feature is that the parent distribution F has disappeared. Whatever process generated the losses, provided the conditions of the theorem hold, the maxima end up in this one family, and so we only need to estimate three parameters rather than guess an entire distribution.

The three parameters of the GEV

• The location parameter μ marks where the distribution of maxima is centred. It measures the central tendency of M_n, not of the parent losses.
• The scale parameter σ measures how widely M_n is dispersed around that centre.
• The tail index ξ (also called the shape parameter) controls how heavy the right-hand tail of the limiting distribution is, and it is the parameter that matters most for risk.

The location and scale parameters play roles similar to a mean and a standard deviation, but they are not the same quantities, and treating them as interchangeable is a common error. For the standardized form (μ = 0, σ = 1) almost all of the probability mass lies between x values of -2 and +6. More generally, most of the mass of a GEV sits between μ – 2σ and μ + 6σ, which shows how lopsided the distribution of a maximum is: it has a short left side and a long right side.

Two practical remarks complete the set-up. First, the same theory handles minima: to model the smallest values of a sample, multiply the data by -1 and apply the maxima results. Second, Figure 1 previews the two ways of harvesting extremes that the rest of the lesson develops. GEV theory works with the block maxima, while the peaks-over-threshold method works with every exceedance of a high level.

The sign of the tail index splits the GEV into three named distributions, and each one corresponds to a different kind of parent tail. Knowing which parent tails lead to which limit is the idea of a domain of attraction: a parent distribution lies in the domain of attraction of a particular GEV case if its sample maxima converge to that case as n grows.

The Frechet distribution is the natural home for financial data. Returns on equities, currencies and commodities are typically fat-tailed, and a power-law tail is exactly what a positive tail index captures. Empirical work tends to find tail index estimates for financial returns that are positive yet below 0.35, which is heavy-tailed but not wildly so. The Gumbel distribution arises when the parent has exponential tails, which is the situation for the normal and lognormal; it is a useful benchmark but, as later sections show, choosing it when the true tail index is positive can be costly. The Weibull distribution arises when the parent has a finite upper end point, so extreme losses cannot exceed some bound, and this is too optimistic for almost any market series.

How the Frechet and Gumbel shapes differ

Both the standardized Frechet and the standardized Gumbel densities are skewed to the right, with a peak just below zero and a long tail stretching towards positive values. The difference is in how far that tail reaches. The Frechet peak sits a little lower and its right tail stays visibly above the Gumbel tail for large x, so the Frechet assigns noticeably more probability to very large values of the maximum. That extra tail probability is the whole point of using it: if large losses come from a power-law world, a Gumbel model will make them look rarer than they are.

A short way to remember the mapping is that extremes drawn from Levy or t-distributions are attracted to the Frechet, while extremes drawn from normal or lognormal parents are attracted to the Gumbel. The parent tail shape decides the destination, and the speed at which the tail decays is what the tail index measures.

To turn the GEV into a risk number we need its quantiles: the value x that the maximum stays below with a chosen cumulative probability p. Setting H(x) = p in Equation 3.1 and taking natural logs of both sides gives the intermediate form below.

Solving each branch for x produces closed-form quantile formulas for the two cases that matter in finance.

Reading down the table, the upper quantile rises as the tail index rises, and the Frechet values move towards their Gumbel counterparts as ξ shrinks towards zero. So Frechet quantiles are sensitive to the tail index, and that sensitivity is the first sign of a theme that runs through the whole subject.

Probabilities here refer to the maxima, not to the parent losses

The p in Equations 3.1 to 3.3 is a probability on the distribution of extreme losses. A 5% quantile of the GEV separates the smallest 5% of sample maxima from the largest 95%. That point lies on the left side of the distribution of maxima, because it is a modest maximum, yet it still lies far out in the right tail of the parent loss distribution, because any sample maximum is a large loss. Confusing the two scales is easy, and the next section shows how to translate between them.

A VaR is a quantile of the parent loss X at confidence level α, so we need a bridge from the distribution of M_n to that of X. Pick some high threshold M_n*. The maximum of n iid losses stays below M_n* only if every one of the n losses stays below it, so the two probabilities are linked by a power of n:

So the α VaR of the parent is the GEV quantile at p = αⁿ. Because ln(αⁿ) = n ln(α), substituting into Equation 3.3 gives the extreme value VaR formulas directly.

The subscript n on each parameter is a reminder that the parameters belong to maxima drawn from blocks of size n. Changing the block size changes the parameters, and keeping the subscript visible prevents mistakes when thinking about the limit as n grows. These formulas are intended for very high confidence levels only; at modest confidence levels the extreme value approximation has no reason to be accurate.

The comparison is clear. At 99.9% the Gumbel gives 2.3021, the Frechet with ξ = 0.2 gives 2.9237 and the Frechet with ξ = 0.3 gives 3.3165. Extreme value VaR, and by extension any extreme value risk measure, moves materially with the tail index, so a good estimate of ξ_n is central to the whole exercise. This holds even when a Gumbel is used, because the Gumbel is only appropriate if ξ_n is statistically indistinguishable from zero.

Scaling extreme value VaR to longer horizons

Longer holding periods can be handled by rescaling the parameters: the location parameter scales roughly in proportion to the horizon h, the dispersion scales roughly with the square root of h, and under certain conditions the tail index does not scale at all. The net effect is that VaR scales as VaR(h) = VaR(1) × h^K. Empirical evidence from Hauksson and co-authors puts K at about 0.45 on average, so the familiar square-root-of-time rule, which sets K = 0.5, overstates the scaling for extreme value distributions.

Example 5 showed that a Gumbel assumption lowers the 99.9% VaR from 4.322 to 3.612 (3.611 on recomputation) for the same location and scale. The choice between the two cases is a risk decision, not a technicality, and there are three broad ways to make it.

Choose by the domain of attraction of the parent

If we are confident about the parent distribution of the losses, the choice follows automatically: use the extreme value distribution whose domain of attraction contains that parent. A risk manager convinced that daily losses are Student t distributed would pick the Frechet, because maxima of t-distributed variables converge to it. Put simply, we choose the limit that the extremes of the parent will actually tend towards. The weakness is that confidence about the parent is exactly what we usually lack.

Choose by testing the tail index

Alternatively, estimate ξ and test whether it differs significantly from zero, using the Gumbel if it does not and the Frechet if it does. This is statistically tidy, but it carries an asymmetric danger. Tail index estimates are noisy, so a test can easily fail to reject ξ = 0 when the true value is positive. Accepting the Gumbel in that situation leads straight to an underestimate of extreme risk measures, which is the more expensive kind of error.

Default to the Frechet

Because the estimated risk measure rises with the tail index, and because model risk is a real concern in this field, the safer course is simply to use the Frechet. If the true tail is exponential, the Frechet fit should return a small ξ and little is lost; if the true tail is a power law, the Gumbel would have been dangerously optimistic. The asymmetry of the costs favours the heavier-tailed choice.

When ξ > 0, a full GEV fit is not the only route to an extreme quantile. The tail probability of a heavy-tailed loss, meaning the chance that a loss exceeds x, behaves like a power of x multiplied by a slowly varying function k(x), one that changes so gradually with x that it is almost constant far out in the tail. In the notation of the source this tail probability is written F(x):

Treating k(x) as a constant k turns this into a pure power law, F(x) = kx^-1/ξ (Equation 3.7). Now compare two tail probabilities: an in-sample probability, which the data can estimate, and a smaller out-of-sample probability, which comes from the VaR confidence level and may lie beyond any observation we have.

Dividing one line by the other cancels the unknown k, which is the trick that makes the method work. Solving for the out-of-sample quantile gives:

The in-sample probability is not known exactly, but it has an obvious empirical proxy. If t of the n observations exceed x_in-sample, then t/n estimates p_in-sample, and the formula becomes:

The recipe needs four ingredients. Pick any convenient in-sample quantile, count how many observations exceed it to get t/n, set the out-of-sample probability equal to one minus the VaR confidence level, and estimate the tail index by a suitable method such as the Hill estimator. The extrapolated quantile then follows in one line.

A related idea, proposed by Diebold and co-authors, takes logs of the power law and estimates the tail by regression. That route remains largely untested and can yield an implausible tail index because nothing in the regression forces a sensible value, so the direct formula above is usually preferred.

Every formula so far needs values for μ, σ and, for the Frechet, ξ. Four families of estimator are available: maximum likelihood (ML), regression, moment-based methods and semi-parametric methods. This section covers the first two, which work on a sample of m block maxima M_n¹, …, M_n^m.

Maximum likelihood estimation

Maximum likelihood picks the parameter values under which the observed maxima are most probable. For the Gumbel the log-likelihood of the m maxima is:

When ξ is not zero the log-likelihood gains the tail index as a third argument:

ML has solid credentials. It is statistically well grounded, the estimators are consistent and asymptotically normal provided ξ_n > -1/2, and hypotheses about the parameters can be tested with likelihood ratio statistics. Its drawbacks are practical. There are no closed-form solutions, so a numerical optimizer and suitable software are needed, the estimates may not be robust, and because the supporting theory is asymptotic, small samples of maxima can cause trouble.

Gumbel’s regression method

A simpler route, due to Gumbel (1958), uses the fact that sorted maxima are order statistics. Arrange the m maxima from lowest to highest. For large n, the expected value of the distribution function at the i-th smallest maximum is i/(1 + m):

Combining this with Equation 3.1 for ξ ≠ 0 and taking logs twice produces a relationship that is linear enough to fit by least squares:

Regressing log[-log(i/(1 + m))] on the bracketed term yields least-squares estimates of μ_n, σ_n and ξ_n. In the Gumbel case the relationship is exactly linear in the maxima:

So a plot of log[-log(i/(1 + m))] against the sorted maxima should be a straight line with slope -1/σ_n and intercept μ_n/σ_n, and the two parameters can be read straight off the fitted line.

Semi-parametric methods target the tail index alone, and they work on the parent losses rather than on block maxima. The most widely used is the Hill estimator. Sort the losses from largest to smallest as X₁, X₂, …, X_n and choose a tail size k. Then:

In words, it is the average log of the k largest losses minus the log of the next one in line, the (k + 1)th, which marks the edge of the tail. Heavier tails spread the largest losses further above that edge, so the average log gap and the estimate both rise.

The Hill estimator is consistent and asymptotically normal, but its behaviour in finite samples is poorly understood and it is sensitive to the choice of k. Even so, many practitioners consider it as good as any alternative. The main competitor is the Pickands estimator, which, unlike the Hill estimator, does not require a positive tail index and is asymptotically normal and weakly consistent under reasonable conditions, though it is less efficient.

Reading a Hill plot

Theory offers little help with k. The common suggestion is to compute the estimator over a range of k values, plot it against k, and look for a region where the line flattens: the level of that plateau is then taken as the tail index. This is the Hill plot, and it tries to use as much of the data as possible, albeit informally.

The trap is that a plateau can be an illusion. In a simulation of 1000 draws from a standard normal, the Hill estimates for k up to 40 wobble at first and then appear to settle in the region of 0.18 to 0.20, which looks like a perfectly reasonable tail index. Extending the plot to k = 200 destroys that impression: the estimates keep moving and never stabilize. Such a graph is called a Hill horror plot, and it gives no guidance on k at all. Normal data might be expected to behave badly here, but horror plots are common with many non-normal distributions too.

A tempting response is to adjust the estimator until the plot behaves. Adding an extra term of -0.0015k to Equation 3.17 turns the same simulated data into a well-behaved “Hill happy plot” that settles around 0.15. The patch has no theoretical basis, and there is no way to know whether the tail index it suggests is any closer to the truth; it merely makes the picture look tidy.

Choosing k by minimizing mean-squared error

With very large samples a more principled method exists. Danielsson and de Vries note that k involves a trade-off between bias and variance. A larger k brings in more data, which lowers the variance of the estimator, but it reaches towards the centre of the distribution and so raises the bias of a tail estimate. A smaller k reduces that bias at the cost of higher variance. They choose k to minimize a mean-squared-error criterion, using a second-order approximation to the tail and a subsample bootstrap to find the size at which bias and variance vanish at the same rate. The procedure needs at least 1500 observations, so it is impractical for small samples, and some writers distrust any automatic rule that ignores softer judgmental information.

A third family of estimators matches empirical moments of the extremes data to their theoretical counterparts. Let m_i be the i-th empirical moment of the maxima. For ξ ≠ 0, an adaptation of results in Embrechts and co-authors gives three equations in the three unknowns, where Γ(·) is the gamma function:

Dividing the third equation by the second cancels both σ/ξ and Γ(1 – ξ), leaving (3m₃ – m₁)/(2m₂ – m₁) = (3^ξ – 1)/(2^ξ – 1), which pins down an estimate of ξ on its own. With ξ in hand, the first two equations rearrange into estimators of scale and location:

The Gumbel versions follow by letting ξ go to zero. The ratio ξ/(2^ξ – 1) tends to 1/ln 2, and (1 – Γ(1 – ξ))/ξ tends to the derivative Γ'(1) = -0.57722, so:

The method is easy to apply but unreliable, because sample moments of second and higher order have poor sampling properties when the data are heavy-tailed. Hosking and co-authors (1985) showed that estimates with much better sampling behaviour come from replacing m_i with probability-weighted moments w_i = E[X(F(X))^i-1] for i = 1, 2, …, or with the more general w_i,r,s = E[X^r(F(X))^i-1(1 – F(X))^s-1] for i, r, s = 1, 2, …. Weighting by the distribution function rather than raising the data to higher powers keeps the influence of a single huge observation in check.

The analysis so far treated the maxima as if they came exactly from a GEV. In practice they are only approximately GEV, meaning the parent lies in the maximum domain of attraction of a GEV case. Handling that properly needs a little more machinery. Take the Frechet case, with ξ = 1/a > 0. The far right tail of the parent then has the form 1 – F(x) = x^-aL(x), where L is a slowly varying function, and membership of the Frechet domain of attraction means that

Applying Equation 3.1 at a large value u = c_nx + d_n gives an approximation to the tail probability, and inverting it gives a quantile estimator at a high confidence level p:

The difficulty is that we usually want p so high that the quantile lies beyond the data, where the empirical tail is zero. The standard remedy is a subsequence trick that replaces n with n/k, so that norming constants can be estimated from a shorter effective sample:

The constants c_n/k and d_n/k come from suitable semi-parametric methods, and ξ comes from the Hill estimator or another tail index approach. Structurally this is the Frechet VaR formula again, with the block size and confidence level combined into the term (n/k)(1 – p).

The second branch of EVT ignores blocks altogether and looks at every loss that exceeds a high threshold. This is the peaks-over-threshold (POT) approach, also called the generalized Pareto approach. It usually needs fewer parameters than the GEV route, and it plays the same role for exceedances over a threshold that GEV theory plays for sample maxima.

Let X be an iid loss with distribution function F(x) and let u be a threshold. The distribution of the excess loss X – u, conditional on the loss being above u, is:

F_u(x) is the probability that a loss overshoots the threshold by no more than x, given that it overshoots at all. The parent F could be normal, lognormal, t or something else, and is usually unknown. The Gnedenko-Pickands-Balkema-deHaan (GPBdH) theorem removes the need to know it: as u rises, F_u converges to a generalized Pareto distribution (GPD):

The GPD has just two parameters: a positive scale parameter β and the tail index ξ, which can be positive, zero or negative and is the same tail index that appears in GEV theory. Of the three cases, the heavy-tailed one with ξ > 0 matters most for market data, and the exponential case with ξ = 0 is the natural benchmark.

The theorem is powerful because the form of the excess-loss distribution is the same, in the limit, almost whatever the parent loss distribution happens to be. If the threshold is high enough, the GPD is the natural model for excess losses.

From excesses back to ordinary losses

Using the GPD requires a threshold u, which fixes N_u, the number of observations above it. Rearranging Equation 3.18 recovers the parent distribution for losses above u:

The only unknown besides the GPD parameters is F(u), the share of observations at or below the threshold. Its natural estimate is the observed proportion (n – N_u)/n. Substituting this and Equation 3.19 into Equation 3.20 gives the tail estimator on which POT risk measures are built:

VaR at confidence level α is the loss x at which the tail estimator in Equation 3.21 equals α. Setting F(x) = α and solving gives:

The structure is easy to read. Start from the threshold u. The term (n/N_u)(1 – α) compares the target tail probability 1 – α with the observed exceedance rate N_u/n; the smaller that ratio, the further beyond u the VaR lies. A larger scale β stretches the distance, and a larger tail index stretches it faster as the confidence level rises.

Expected shortfall is VaR plus the average amount by which losses beyond VaR exceed it. For the GPD that mean excess is linear in the level, which gives a closed form provided ξ < 1:

The gap between ES and VaR widens as the confidence level rises. That is a direct consequence of the positive tail index: under the GPD the expected excess beyond a level grows with the level, so the further out we go, the larger the average overshoot.

The threshold is the weak spot of POT. There is no objective rule for setting it, so the choice is always partly arbitrary and judgmental. The trade-off mirrors the choice of k in the Hill estimator. A high threshold makes the GPBdH approximation more accurate, because the theorem is a statement about the limit as u rises. Push it too high, however, and N_u becomes so small that the parameter estimates are unreliable. Lower the threshold and the opposite happens: more exceedances, more precision, but a poorer approximation.

Using the mean-excess function

The same informal device used for k can help here. Plot the mean-excess function (MEF), the average excess of losses over a candidate threshold, against that threshold. For a GPD the mean excess is linear in the threshold, and the advice is to pick a threshold beyond which the MEF settles into a stable pattern, ideally becoming flat. As with the Hill plot, this works when the picture is clean and gives little guidance when it is not.

Maximum likelihood for the GPD

Once u is fixed, ξ and β can be estimated by maximum likelihood or other methods, and ML is usually regarded as the most reliable. Writing X_i for the m excess losses, the log-likelihood is:

Provided ξ > -0.5, the ML estimators are asymptotically normal and so reasonably well behaved. Moment-based estimators are also available. From the first two sample moments of the excesses, m₁ and m₂:

As with the GEV, these moment estimators can be unreliable, and probability-weighted or ML estimators are generally preferred.

The GEV and POT approaches are two expressions of one theory. The GEV models the distribution of the extremes themselves, the largest value in each block, while POT models the distribution of exceedances over a high threshold. Both rest on the same tail index, and in principle neither dominates. In practice there can be reasons to prefer one.

The GEV usually carries one more parameter than the GPD, and the block maxima method, the most common GEV implementation and the one assumed in the earlier sections, can throw away useful information: a block containing two or three severe losses contributes only its largest, as blocks one and four in Figure 1 show. Both points count against the GEV. On the other side, POT forces the analyst to pick a threshold, a problem that never arises with block maxima. Which approach is more natural may also depend on the data available; with limited data, one may simply fit more comfortably than the other.

The sensible conclusion is that either approach is generally reasonable, and the choice should follow whichever suits the problem at hand.

Everything above is unconditional: the methods are applied directly to the loss variable X without adjusting for any dynamics. Unconditional EVT suits long forecasting horizons. Over short horizons, however, losses often have a structure we can model, and ignoring it wastes information.

Conditional or dynamic EVT

Suppose returns follow a GARCH process. Then the raw returns are not iid, but the innovations driving them may be. Conditional EVT applies the extreme value machinery to those innovations rather than to the returns. McNeil and Frey set out a two-step procedure. First, estimate a GARCH-type model by a suitable econometric method and extract its residuals, which should be close to iid; the fitted model also forecasts the next period’s location and scale, μ_t+1 and σ_t+1. Second, fit an EV model to the residuals and combine its tail quantile with the GARCH forecasts, so that the final VaR reflects both the volatility dynamics and the residual tail.

Why clustering matters

Most financial returns show time dependence, usually in the form of clustering: large moves arrive in bunches. This matters for three reasons. It breaks the iid premise on which the earlier results rest, and the statistical consequences are not well understood. There is evidence that dependence can make estimators perform very poorly. It also changes how results should be read. A VaR expected to be breached once every so often on average may, under clustering, be breached several times in one turbulent month and not at all for long stretches, so the number of breaches in a given period depends on how strongly they cluster.

Three ways to deal with dependence

• Block maxima. The simplest and most common fix is to apply the GEV to block maxima. Maxima are less clustered than the raw data, and the longer the block, the weaker the dependence, so long enough blocks can remove it entirely. The price is lost efficiency, since non-maximal extremes are discarded, and a new bandwidth problem: there is no clear rule for block length, much as there is none for k.
• Model the conditional distribution. Fit a conditional volatility model such as GARCH first and estimate the tail index on the standardized data. The deterministic part of the model absorbs the time dependence, and the remaining random process can be treated as independent.
• The extremal index. A more advanced and more difficult option estimates the extremal index, a measure of clustering, and uses it to adjust the quantiles.

Portfolios hold many risk factors, so the joint behaviour of extremes matters as much as each tail on its own. Multivariate extreme value theory (MEVT) models the tails of multivariate distributions in a theoretically consistent way, and its central question is how to represent the dependence between extreme events.

Why correlation is not enough

In central-value statistics the central limit theorem justifies a normal or, more broadly, an elliptical distribution. Within that family, linear correlations together with variances, or equivalently the covariance matrix, fully describe the joint distribution, which is why correlation is so prominent in everyday risk work. Once we move outside the elliptical world, correlation no longer captures the dependence structure. MEVT shows that the limiting joint distribution of multivariate extremes belongs to a family of extreme value copula functions, so multivariate extreme dependence is modelled by choosing one of these EV copulas rather than by estimating correlations.

The curse of dimensionality

In principle an EV copula can have any number of dimensions. In practice the data run out quickly. Suppose each variable has a one-in-100 extreme event and the variables are independent. A joint extreme of two variables would then be expected only once in 100 squared, or 10,000, observations, and a joint extreme of three variables only once in 100 cubed, or 1,000,000. Each added dimension makes joint extremes rarer while adding parameters to estimate, which is the curse of dimensionality, and it sets a practical limit on how many variables MEVT can handle.

Tail dependence is the real issue

It would be wrong to conclude that joint extremes are too rare to worry about. That arithmetic assumed independence, and extremes are not independent in general: how often they coincide depends on their joint distribution, specifically on the tail dependence of that distribution. This is why tail dependence is the focus of MEVT. Experience confirms it: a major earthquake can set off a tsunami or a market crash, and financial disasters frequently arrive together. So a risk manager needs at least an awareness of multivariate extreme risk, even when a full multivariate model is out of reach.

EVT gives a purpose-built way to estimate extreme probabilities and quantiles. It is intuitive, it is relatively easy to apply in its basic forms, it tells the analyst what to estimate and how, and it has a good track record. Just as important is what it rules out: distributions justified by central limit reasoning, above all the normal, should not be used for extreme value estimation, because they contradict what extreme value theorems say about tails.

The approach still has clear limits, and they deserve equal weight.

• Scarce data. Extreme observations are few by definition, so EV estimates are much less precise than estimates of central quantiles and carry wide confidence intervals. This is not a flaw in EVT but an unavoidable result of having little data.
• Model risk. EV estimation rests on assumptions whose truth is hard to verify, and the results are often very sensitive to them. Ancillary choices, such as the threshold, the block length or k, have no easy answers and call for a good deal of subjective judgment. Because of this, confidence intervals for EV risk measures should always be estimated, and the measures should be subjected to stress testing.
• Small-sample problems. Most of the theory is asymptotic, and with so few data EV estimates can be distorted by small-sample effects, biases and non-linearities.

McNeil makes the balance explicit: working in the tail with limited data means high uncertainty and large confidence intervals, yet the theoretically supported methods of EVT still beat ad hoc alternatives and give the most honest measure of that uncertainty. Diebold and co-authors put the same point from the other side, warning that EVT should not be asked to deliver more than it can. Its lasting contribution is to draw smooth, theory-guided curves through the extreme tails of empirical survival functions. Used with that understanding, it has a valuable but bounded role in risk measurement.