Loss quantiles and tail averages

What you will learn

Compute an explicitly defined loss quantile and worst-quarter expected shortfall in an eight-case model, distinguishing strict exceedances from boundary mass.

Prerequisites: Expected value as a weighted average · Variance and standard deviation · Unequal outcome weights

Define the sign and quantile convention

Let L be signed loss: a positive number is a loss, and a negative number is a gain. A lower α-quantile is the smallest loss threshold x for which P(L≤x) is at least α. This convention matters when a finite distribution places positive mass at individual values. A quantile is a threshold, not a maximum possible loss or an average loss beyond it.

In eight equally likely cases, the lower 75% quantile is the sixth increasing ordered loss. For losses 0,0,0,0,0,0,10,30, it is zero. Six of eight cases have loss at most zero; the remaining cases can still lose 10 or 30. Calling the threshold a zero-loss guarantee would confuse probability coverage with a bound on the support.

Average a fixed amount of worst-case mass

At the 75% level, expected shortfall here means the average loss in exactly the worst 25% probability mass. Eight equal cases make this the mean of the two greatest losses. In the first model, (10+30)/2=20. We use an exact supplied finite population, not a historical estimate or a claim that future losses follow it.

A boundary tie needs care. Change the losses to −40,0,0,0,0,0,0,80. Mean loss is still 5 and the lower 75% quantile is still zero, but the worst quarter consists of the 80-loss case plus one of the zero-loss cases. Its average is 40. Tied zero cases are interchangeable for this calculation; selecting all of them would expand the tail mass far beyond 25%.

Strict exceedances and boundary-inclusive sets differ

In the second model, only 1/8 of cases lie strictly above the zero quantile, but 7/8 lie at or above it. Averaging strictly above would give 80, while averaging all cases at or above gives 80/7. Neither equals the fixed-mass worst-quarter mean of 40. Discrete atoms make the shortcut “average everything beyond the threshold” ambiguous unless its boundary convention is specified.

These metrics summarize different aspects of a stated distribution. Equal means and quantiles do not imply equal tail severity. An expected shortfall is itself an average, so an individual loss can exceed it; it is not a cap or a prediction of what must happen next. Real estimation additionally needs justified probabilities, a horizon, appropriate data and model validation, which the eight-case exercise does not provide.

Hands-on: Order signed losses and construct fixed tail mass

Place eight physical cases, select exactly the worst quarter and distinguish a lower quantile, tail mean and strict versus boundary-inclusive probabilities.

Reading tail workshop…

Common mistakes

• Treating negative loss as a worse loss rather than a gain.
• Reading a quantile as a maximum possible loss.
• Averaging all boundary-tied cases instead of the required tail mass.
• Confusing exact supplied-population arithmetic with validated future coverage.

Keep these ideas

• Declare the sign and lower-quantile convention.
• Order the complete probability distribution.
• Average exactly the stated worst tail mass, with boundary ties handled explicitly.
• Separate a threshold, a tail average and a maximum.

Further reading

• Acerbi and Tasche — Expected Shortfall (2001 preprint), pages 4–5 ↗

  Author preprint, pages4–5: threshold versus worst-tail average and boundary ties. Source uses returns and a lower-tail sign convention; our loss L reverses the sign and uses an upper worst25% tail. Original exact eight-case populations are not an estimator of future loss. No whole-paper review claimed. Checked 2026-10-21.