Percentile Calculator

Percentile and percentage are unrelated

Scoring in the 90th percentile does not mean answering 90 percent correctly. It means scoring at or above 90 percent of the other people who took it. A raw score of 45 percent can be a 95th-percentile result on a hard exam. Percentile is a rank within a distribution; percentage is a proportion of a total, and conflating them is the most common misreading of test results.

There is more than one correct formula

Statistical software implements several percentile definitions — the standard lists nine — and they disagree, particularly on small datasets. The main split is whether to interpolate between adjacent values or to select an actual data point, and whether the rank is computed as n times p or as n plus one times p. Excel's older PERCENTILE, its newer PERCENTILE.EXC, R's default and NumPy's default do not all return the same figure. Reporting a percentile without saying which method produced it is incomplete, and comparing percentiles from two different tools is unsafe.

Quartiles, deciles and the IQR

Quartiles split data at the 25th, 50th and 75th percentiles; deciles split at every tenth. The interquartile range, the gap between the 25th and 75th, measures spread while ignoring the extremes entirely, which makes it far more robust than standard deviation on skewed data. The common outlier rule flags anything more than 1.5 times the IQR beyond either quartile.

The median is the 50th percentile

And it beats the mean whenever the distribution is skewed. Income, house prices, response times and file sizes all have long right tails, where the mean sits well above what most observations look like. Any headline that reports an average of such data is describing the tail more than the typical case.

Latency percentiles are where this matters most

Average response time hides the experience of the slowest requests. Engineering teams track p95, p99 and p99.9 because those are the users having a bad time, and because a system with a 100 ms mean and a 4 second p99 is failing one request in a hundred. On a page making twenty backend calls, a p99 problem touches roughly one page load in five, which is why tail latency dominates perceived quality.

Percentiles do not average

The p99 of two servers is not the mean of their two p99 values, and combining percentiles across time buckets or shards requires the underlying distribution, not the summary figures. Averaging percentiles is a common monitoring error that produces numbers looking reasonable and meaning nothing.