Tail Value-at-Risk (Expected Shortfall) as a Tail Summary

Description

Summarise the bad tail of a distribution by its high percentile and by the average of all outcomes beyond that percentile.

UX and performance reporting often stops at a high percentile, such as p95 load time. That is Value-at-Risk: it says where the tail starts but nothing about how bad it is beyond that point. Tail Value-at-Risk, also called expected shortfall or conditional tail expectation, is the average of all outcomes beyond the percentile, equal to the percentile plus the mean excess above it, so it reflects the whole tail. Klugman and colleagues note these measures are useful for any random variable. TVaR satisfies coherence properties that the percentile alone does not: in particular, combining two risks never yields more than the sum of their separate TVaRs. Empirically, with sorted data, the percentile is the k-th value with k = floor(pn) + 1 and TVaR is the mean of the values from k upward; its naive standard error is too small, so the Manistre-Hancock variance is used.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. Choose the tail level p (for example 0.95 or 0.99).
  2. Sort the observations or simulated outcomes.
  3. Set k = floor(pn) + 1; the percentile (VaR) is the k-th smallest value.
  4. Compute TVaR as the mean of the observations from the k-th upward.
  5. Estimate the TVaR variance with the Manistre-Hancock formula (or a bootstrap) and form a confidence interval.
  6. Report percentile and TVaR together, and track both over time or compare them across designs.

Sources

  • Loss Models: From Data to Decisions (Klugman et al., 2012)
  • Preparing for the Worst: Incorporating Downside Risk in Stock Market Investments (Vinod & Reagle, 2005)

Also known as

  • Tail Value-at-Risk
  • TVaR
  • Expected shortfall
  • Conditional tail expectation
  • CTE
  • CVaR
  • Value-at-Risk
  • Tail-Value-at-Risk (expected shortfall) and Value-at-Risk as tail summaries

Categories

  • Analytics & Behaviour
  • Statistics & Modelling