Statistical Methods
Distinguish an Expectile from a Percentile
Define the task
An expectile summarizes location using asymmetric squared loss. Its level is not the fraction of observations below the estimate, so an 0.8 expectile is not an 80th percentile. Select the target according to the question and consequences of positive versus negative deviations. The browser tool uses equal observation weights and a level strictly between zero and one. It supplies a descriptive estimate without a confidence interval or model for dependence.
Carry out the check
Find a location e where alpha times the sum of positive deviations above e equals (1−alpha) times the sum of deviations below e. The level 0.5 gives the arithmetic mean. Larger levels give greater weight to the upper side, and extreme observations can still exert substantial influence because the underlying loss is squared. The tool brackets the result between the minimum and maximum and solves the balance equation numerically. Keep units and the chosen level in the report.
Worked hypothetical example
For hypothetical values (0,2,10) at alpha=0.8, e=7 balances 0.8×(10−7)=2.4 against 0.2×[(7−0)+(7−2)]=2.4. Only one of three observations lies above seven; the level is plainly not a percentile count. At level 0.5 the estimate is the mean, four. Constant observations return that constant at every accepted level. Compare alternative levels only when the sensitivity analysis has a clear purpose, and do not label the most favorable level as prespecified after inspection.
Method checked 19 November 2026 against Primary method documentation: expectile.