Percentiles

A percentile estimate is a probability and value pulled from a distribution. Some examples:

The 90% percentile of customer losses is $1M.

This communicates a 90% chance that future customer losses will be $1M or less.

It also communicates a 10% chance that losses will exceed $1 million.

The upside: A percentile estimate is very simple to gather and communicate.

The downside: It won’t communicate how the risk behaves.

Stated differently: The possible outcomes might be normally distributed, exponential, uniform, or something else. This sort of intuition is not communicated from from a single percentile estimate.

Some more examples of what a percentile estimate can sound like:

• We expect 95% of fraud claims to lie under $1M.
• The median (50% percentile) customer account value is $1K.
• 75% of our incidents last under 3 days to investigate.

Elicitation is quick and often understood. Though, a sole percentile estimate won’t disclose other information about a risk (extreme values or underlying distributions).

For instance, let’s say the 90% height in a group of people happens to be 6 foot 3. Knowing that height is normally distributed, we don’t expect the remaining 10% of people to include any 20 foot giants. Rather, we expect the remaining 10% of people to be able to walk through a doorway.

However, a 90% percentile estimate of regualtory settlements may be entirely different. Assume 90% of settlements are expected to be under $200 million for a business. Within that remaining 10% of possible settlements may be some extreme values in the multiple billions.

The percentile itself does not communicate distributions of risk. Still, a percentile estimate is quick and useful especially if an audience is familiar with the underlying behavior of the risk.