Percentile
A percentile is a simple way to say “how a value compares to the rest of the group.” Instead of focusing on the average, it tells you what fraction of the data falls below a particular point.
What a percentile means
The p-th percentile is a value such that about p% of observations are at or below it (and about 100 − p% are above it). For example, the 90th percentile is the cutoff where roughly 90% of the data is lower and 10% is higher. Percentiles are closely related to quantiles: the 25th, 50th, and 75th percentiles are the first quartile (Q1), median (Q2), and third quartile (Q3).
Why it’s useful (especially with skewed data)
Averages can be misleading when a dataset has extreme values (outliers) or is lopsided. Percentiles are more robust because they care about rank order, not the size of the extremes. That’s why dashboards often report things like p50 (typical), p90 (high-but-not-maximum), and p99 (near-worst-case).
Practical examples
- Exam scores: being in the 80th percentile means you scored higher than about 80% of students.
- House prices: the 75th percentile price marks the point above which the top 25% most expensive homes sit.
- Model latency: “p95 latency = 200 ms” means 95% of requests finish within 200 ms; the slowest 5% take longer.
Why it matters in AI/ML workflows
Percentiles show up when you:
- Set thresholds (e.g., flag the top 1% most suspicious transactions).
- Monitor drift by tracking changes in feature percentiles over time.
- Summarize performance beyond the mean (common in ML systems and A/B testing).
Percentile is a value below which a given percentage of observations in a dataset fall; the pth percentile is the point where p% of data are at or below it. Percentiles summarize a distribution’s location and spread and are robust to outliers, making them useful for monitoring model inputs, detecting drift, and setting thresholds. Example: the 95th percentile of prediction latency helps define an SLO for worst-case user experience.
Imagine a long line of students sorted by height, from shortest to tallest. If you’re at the point where 90% of students are shorter than you, you’re at the 90th percentile. A percentile is simply a “position in the line” that tells you what percentage of the data falls below a certain value.
In statistics and AI/ML, percentiles help describe how data is spread out. For example, the 50th percentile is the median (the middle value). Percentiles are also used to spot unusually high or low values, like detecting outliers in sensor readings or user activity.