Quantiles work by reversing the usual probability question. Instead of asking, “What probability lies below this value?”, a quantile asks, “What value marks this cumulative probability level?” The median is the 50th percentile, the lower quartile is near the 25th percentile, and a 95th percentile is a threshold below which roughly 95% of the distribution lies under the chosen definition. Quantiles preserve ordering and probability position, which makes them robust and decision-relevant, especially in skewed or heavy-tailed distributions. But they deliberately discard much of the distance information between observations and their exact numerical magnitudes away from the selected rank.
Suppose a delivery company says its average delivery time is 24 minutes.
That sounds useful.
But a customer may care about another question:
How long must I wait before 95% of deliveries have already arrived?
That is a quantile question.
NIST defines percentiles through ordered observations and explicitly notes that multiple interpolation conventions are used in practice. That detail matters: for small samples, a “90th percentile” is not always a uniquely determined sample number unless the quantile convention is stated.
Quick Read
DISTRIBUTION / ORDERED DATA → CUMULATIVE PROBABILITY p → INVERT CDF → p-QUANTILE → MEDIAN / QUARTILE / PERCENTILE → SAMPLE ESTIMATION → INTERPOLATION → UNCERTAINTY → TAIL THRESHOLD → DECISION / RISK / PREDICTION
1. Quantiles Start With Ordering
Take observations and sort them from smallest to largest.
Order statistics are the sorted values.
NIST defines the rank of an observation by its position in this ordered sequence.
Quantiles use those ranks to locate probability positions.
2. The Population Quantile Is an Inverse-CDF Idea
A common mathematical definition is:
Q(p)=inf{x:F(x)≥p}.
Probability p goes in.
A value threshold comes out.
This generalised inverse works even when the CDF has jumps.
3. The Median Is the Central Quantile
The median is a 0.5 quantile under a convention.
It divides probability so that at least half lies at or below and at least half lies at or above under standard definitions.
Unlike the mean, it depends mostly on order rather than magnitude.
4. Quartiles Divide the Distribution Into Four Probability Regions
Q1 marks roughly the 25th percentile.
Q2 is the median.
Q3 marks roughly the 75th percentile.
The interquartile range Q3−Q1 measures the width of the middle 50%.
5. Percentiles Divide Probability Into Hundredths
The 90th percentile is a threshold associated with cumulative probability 0.90.
Under an appropriate continuous distribution, about 90% of values lie below it and 10% above.
Percentile rank and percentile value should not be confused.
6. Percentile Rank Answers the Reverse Question
Quantile question:
“What score is the 90th percentile?”
Percentile-rank question:
“What percentage of scores lie at or below 82?”
One maps probability to value.
The other maps value to cumulative probability.
7. Quantiles Are Not Means
A distribution can have mean 100 and median 60.
That happens when large upper-tail values pull the mean upward.
Quantiles preserve probability position; means preserve weighted magnitude.
8. Quantiles Are Robust to Extreme Magnitudes
Move the largest observation from 100 to one billion.
The sample median may not move at all.
Upper extreme values change magnitude dramatically without changing the middle rank.
This is why medians and IQRs are resistant summaries.
9. Robustness Comes From Information Loss
The median resists extremes because it ignores most of their magnitude.
That is not free.
If extreme magnitude is the scientific target, rank-based robustness can discard exactly what matters.
10. Quantiles Transform Naturally Under Monotone Functions
For a strictly increasing transformation g:
Q_{g(X)}(p)=g(Q_X(p))
under standard regularity and quantile conventions.
Order is preserved, so percentile position survives transformation.
11. Means Do Not Transform This Cleanly
For nonlinear g, E[g(X)] generally differs from g(E[X]).
Quantiles have a special compatibility with monotone transformations because rank order survives.
12. Sample Quantiles Require a Convention
A population CDF defines a clean conceptual quantile.
A finite sample often places the desired percentile between observed ranks.
Software then needs an interpolation rule.
NIST explicitly notes that there is no universally accepted interpolation convention and documents several methods in practical use.
13. R6, R7 and R8 Can Give Slightly Different Answers
NIST describes several quantile estimation rules, including methods corresponding to R’s R6, R7 and R8 conventions.
For large samples the differences are usually small.
For small samples and extreme percentiles they can be noticeable.
A reproducible analysis should state the method when it matters.
14. Extreme Quantiles Need Enormous Samples
The median uses information near the middle of the data.
The 99.99th percentile lives in a region containing roughly one observation in ten thousand on average.
Direct empirical estimation of extreme quantiles therefore becomes unstable quickly.
15. Tail Models Can Extrapolate Beyond Observed Quantiles
Extreme-value theory models tail behaviour using specialised limit distributions.
This can estimate rare quantiles beyond observed ranks.
The price is stronger tail assumptions.
Extrapolation becomes model-dependent precisely where data are sparse.
16. Value at Risk Is a Quantile
In financial risk, Value at Risk at confidence level p is essentially a loss quantile under a sign convention.
It asks for a threshold that losses exceed only with tail probability 1−p.
VaR says where the tail begins.
It does not say how bad losses become after crossing that threshold.
17. Expected Shortfall Complements VaR
Expected shortfall averages losses in the tail beyond a selected quantile under suitable definitions.
VaR preserves the threshold.
Expected shortfall preserves average severity beyond it.
They answer different risk questions.
18. Service-Level Agreements Are Quantile Problems
“95% of requests respond within 200 ms” is a percentile statement.
Mean latency could be excellent while the 99th percentile is disastrous.
High-quantile monitoring exposes tail user experience that averages hide.
19. Queueing Systems Often Need Several Quantiles
Median wait describes a typical user.
90th percentile describes a bad but common wait.
99.9th percentile describes rare operational stress.
No single quantile owns the entire service experience.
20. Quantile Regression Models Conditional Quantiles
Ordinary least squares often models E(Y|X).
Quantile regression can model Q_Y(p|X), such as the conditional median or 90th percentile.
This reveals how predictors affect different parts of the outcome distribution.
21. A Predictor Can Affect the Tail Without Moving the Median Much
A new process may leave median manufacturing time unchanged and greatly reduce the 95th percentile.
Mean regression can dilute that operational improvement.
Quantile regression can place the question at the part of the distribution where the decision lives.
22. The Check Loss Makes Quantiles Optimal
Quantile regression minimises an asymmetric absolute-loss function.
The p-quantile is optimal when underprediction and overprediction are weighted according to p and 1−p.
This gives quantiles a direct decision-theoretic meaning.
23. Inventory Decisions Are Quantile Decisions
If stockouts are more expensive than leftover inventory, the optimal stock level is above the median forecast under a standard newsvendor model.
The exact quantile depends on the relative costs.
The decision does not ask for expected demand alone.
24. Prediction Intervals Use Quantiles of Predictive Distributions
A central 90% predictive interval can be formed from the 5th and 95th predictive quantiles under one convention.
Unlike a confidence interval for a parameter, a prediction interval targets a future observation or outcome.
Quantiles convert predictive probability into actionable bounds.
25. Equal-Tailed and Highest-Density Intervals Are Different
An equal-tailed 95% interval uses the 2.5th and 97.5th quantiles.
A highest-density interval seeks a high-probability region of minimum width under a posterior density and can have different endpoints in skewed distributions.
Quantile endpoints therefore encode one interval philosophy, not every possible interval definition.
26. Tolerance Intervals Are About Population Coverage
NIST distinguishes tolerance intervals from ordinary confidence intervals.
A tolerance interval aims to cover a specified proportion of the population with a specified confidence level.
Population percentiles therefore appear directly in quality and reliability work.
27. Quantile Uncertainty Is Sampling Uncertainty
A sample median is itself a statistic.
Another sample produces another median.
Quantile estimates therefore have sampling distributions and standard errors.
See How Sampling Distributions Work.
28. Quantile Standard Errors Depend on Density Near the Quantile
A quantile is easier to estimate when probability mass is dense near its location.
If the distribution is very flat near the target percentile, small probability uncertainty translates into large value uncertainty.
This is why extreme quantiles and sparse regions are difficult.
29. Bootstrap Methods Can Estimate Quantile Uncertainty
Resample the observed data with replacement.
Compute the target quantile in each bootstrap sample.
The bootstrap distribution approximates repeated-sample quantile variability under assumptions.
Extreme quantiles can challenge ordinary bootstrap accuracy when tails are sparse.
30. Discrete Distributions Can Have Non-Unique Quantiles
If the CDF jumps across p, many values can satisfy informal percentile descriptions depending on convention.
The generalised inverse provides one standard mathematical resolution.
Software and textbooks may differ at discontinuities.
31. Ties Matter in Empirical Quantiles
Exam scores and rating scales often contain many equal values.
A single score can therefore span a range of percentile ranks.
Reporting “the 80th percentile” as though ranks were continuously distinct can create false precision.
32. Percentile-Based Grading Is Relative, Not Absolute
A student at the 90th percentile performed above most of the reference group.
That does not say whether the student mastered a curriculum standard.
Criterion attainment and percentile rank answer different questions.
33. Reference Populations Define Percentile Meaning
90th percentile among Primary 6 students.
90th percentile among all adults.
90th percentile among elite mathematics competitors.
The same numerical score can have different percentile positions across reference populations.
34. Percentiles Can Drift Over Time
If the reference population changes, percentile ranks change even when individual ability does not.
Norm-referenced scores therefore require dated and appropriate norm populations.
35. Quantiles Are Essential for Calibration
A probabilistic forecast claiming a 90% upper bound should be exceeded about 10% of the time under calibration and stable conditions.
Quantile calibration checks whether forecast probability positions match observed frequencies.
36. Pinball Loss Evaluates Quantile Forecasts
The asymmetric check or pinball loss penalises underprediction and overprediction differently according to the target quantile.
It is a proper scoring rule for conditional quantiles under suitable conditions.
A 90th-percentile forecast should not be evaluated with ordinary squared error alone.
37. Multiple Quantile Forecasts Can Cross
A separately fitted 90th quantile can accidentally fall below a fitted 80th quantile.
That violates the required ordering of a coherent CDF.
Joint modelling or monotonicity constraints can prevent quantile crossing.
38. Quantile Functions Can Generate Random Variables
If U is Uniform(0,1), then under suitable definitions:
X=Q(U)
has the target distribution.
This inverse-transform method turns uniform random numbers into samples from other distributions.
39. Q-Q Plots Compare Quantile Functions
Take corresponding probability positions from empirical and theoretical distributions.
Plot one quantile against the other.
Systematic curvature reveals shape mismatch.
Tail departures become visible at the ends.
40. Probability Plots Are Distribution-Diagnostic Tools
NIST’s distribution-fitting framework uses probability plots and PPCC methods to assess whether a candidate family adequately represents data.
Quantiles become a visual bridge between observed ranks and theoretical probability structure.
41. Median Absolute Deviation Uses a Quantile Twice
Find the median.
Compute absolute deviations from it.
Take the median of those deviations.
MAD is therefore a robust scale statistic built from quantile operations rather than squared magnitudes.
42. Boxplots Are Quantile Compressions
A boxplot usually displays Q1, median, Q3 and whisker rules derived from the IQR.
It compresses a distribution into rank-based landmarks.
Multimodality and fine shape can disappear completely.
43. Two Distributions Can Share Quartiles and Differ Elsewhere
Matching Q1, median and Q3 does not determine the tails or internal shape between quartiles.
A boxplot can therefore make very different distributions look similar.
Quantile summaries are useful because they compress—not because they preserve everything.
44. Conditional Quantiles Can Reveal Inequality
A policy may raise the median income slightly while raising the 90th percentile substantially and leaving the 10th percentile unchanged.
Mean effects can hide this distributional heterogeneity.
Quantile treatment-effect methods investigate distributional differences under additional causal assumptions.
45. Quantile Treatment Effects Need Careful Causal Interpretation
The difference between treatment and control outcome quantiles is not generally the effect on the person who would sit at that same rank under both treatments.
Rank preservation is an additional assumption.
Distributional causal effects and individual causal effects are distinct.
46. Quantiles Are Useful When Moments Do Not Exist
The Cauchy distribution has no finite mean or variance.
Its median and other quantiles are perfectly well defined.
Rank-based summaries can survive heavy tails that destroy moment-based summaries.
47. Quantiles Can Still Be Hard in Very Heavy Tails
A median may be stable.
An extreme 99.999th percentile may remain nearly impossible to estimate from limited data.
Robustness of central quantiles does not imply easy estimation of extreme ones.
48. What Quantiles Preserve
- ordering;
- cumulative probability position;
- tail thresholds;
- monotone-transformation structure;
- robust central location for the median;
- decision points under asymmetric loss.
49. What Quantiles Discard
- most distance information away from the selected rank;
- the magnitude of values beyond a threshold;
- the full density between quantiles;
- causal mechanism;
- mean contribution of extreme magnitudes;
- multimodality unless many quantiles are inspected.
50. The Failure Created by Forgetting What Quantiles Discard
Two loss distributions have the same 99% VaR.
Beyond that threshold, one maxes out at a modest additional loss.
The other has a catastrophic unbounded tail.
The quantile preserved the boundary and discarded severity beyond it.
51. The Hostile Test: Percentile Used Without a Reference Population
A report says a child is “at the 85th percentile”.
No age group, cohort, country or norm year is stated.
The percentile has no stable interpretive frame.
52. The Second Hostile Test: Quantile Algorithm Changes the Small-Sample Result
Ten observations are used to report the 90th percentile.
Two software packages return different values because they interpolate differently.
Neither is necessarily defective.
The convention was hidden.
53. The Third Hostile Test: Median Used When Magnitude Matters
A reliability team studies catastrophic repair cost.
It reports only the median because it is robust.
The rare but enormous losses were exactly the safety problem.
Robustness discarded the decision target.
54. The Fourth Hostile Test: 95th Percentile Called a 95% Guarantee
A model estimates a 95th percentile under historical conditions.
The report says future values “will stay below this 95% of the time” as though model and environment were known perfectly.
Parameter uncertainty and distribution shift have been erased.
55. Primary School: Quantiles Begin as “Put the Numbers in Order and Find the Position”
Sort nine scores.
Find the middle score.
Then split the lower and upper halves.
The child learns that some summaries depend on position rather than arithmetic magnitude.
A quantile tells us where a chosen fraction of the distribution has accumulated.
56. Secondary School: Compare Mean and Median Under One Extreme Value
Start with 10, 11, 12, 13, 14.
Replace 14 with 1,000.
The mean changes enormously.
The median remains 12.
The resistance of rank-based summaries becomes visible.
57. JC and University: Quantiles Become Inverse Probability Coordinates
At higher levels, learners should reconstruct:
- order statistics;
- CDF inversion;
- generalised inverse definitions;
- sample interpolation conventions;
- asymptotic quantile variance;
- bootstrap uncertainty;
- extreme quantiles;
- quantile regression;
- pinball loss;
- VaR and expected shortfall;
- prediction quantiles;
- quantile calibration;
- Q-Q plots;
- transformation properties.
58. Where Quantiles Fit in the eduKateSG “How Works” Landscape
- How Probability Distributions Work — owns the CDF quantiles invert.
- How Expected Value Works — owns magnitude-weighted centres; quantiles own probability positions.
- How Robust Statistics Work — uses medians, IQR and MAD as resistant summaries.
- How Sampling Distributions Work — owns repeated-sample uncertainty of sample quantiles.
- How Regression Works — owns conditional modelling broadly; quantile regression targets conditional percentiles.
- How X Works Hub — the wider mechanism map.
Quantiles own one precise canonical job: map cumulative probability positions into value thresholds, preserving order and tail location while deliberately discarding much of the magnitude structure outside the selected ranks.
59. What This Article Does Not Claim
- A percentile is not the same as a percentage score.
- A percentile requires a reference distribution or population.
- Sample quantiles can depend on interpolation convention.
- The median is robust because it discards extreme magnitude information.
- A VaR quantile does not describe the severity of losses beyond the threshold.
- Extreme quantiles can be highly uncertain even in apparently large datasets.
- Quantile regression targets conditional quantiles, not conditional means.
- Differences in marginal quantiles are not automatically individual causal effects.
- Equal-tailed intervals are not the only probability intervals.
- Quantiles can remain defined even when moments fail to exist.
60. A Compact Quantile Audit
- What distribution or reference population defines the quantile?
- What probability level p is being used?
- Is this a population or sample quantile?
- What quantile convention or interpolation method is used?
- Are ties or discreteness important?
- Is the target central or extreme?
- How many observations actually inform the tail?
- What is the sampling uncertainty of the quantile estimate?
- Would bootstrap or tail modelling help?
- Is the percentile being confused with a percentage score?
- Does the reference population match the intended interpretation?
- Has the reference distribution shifted over time?
- Would a mean or expected loss answer a different question?
- Does asymmetric loss make a quantile the optimal decision point?
- Is quantile regression needed?
- Are multiple fitted quantiles ordered coherently?
- Does the quantile ignore severity beyond a threshold?
- Would expected shortfall or another tail metric be needed?
- Could a monotone transformation simplify interpretation?
- Is the decision really about rank position or about magnitude?
61. Frequently Asked Questions
What is a quantile?
A quantile is a value associated with a cumulative probability level. A common formal definition is Q(p)=inf{x:F(x)≥p}. The median is a 0.5 quantile.
What is the difference between a percentile and a percentage?
A percentage score describes a fraction of available points or quantity. A percentile describes relative position in a reference distribution.
Why can software give different percentiles?
Finite samples often require interpolation between ordered observations. Several accepted quantile conventions exist, and they can return slightly different values, especially for small samples or extreme percentiles.
Why are quantiles useful?
They describe probability positions directly, remain resistant to extreme magnitude for central quantiles, support tail thresholds, prediction intervals and service levels, and are optimal under certain asymmetric loss functions.
62. Authoritative Research Corridor
- NIST/SEMATECH — Percentiles, Order Statistics and Quantile Methods
- NIST/SEMATECH — Probability Distributions
- Penn State STAT 414 — Probability Theory
- eduKateSG — How Probability Distributions Work
- eduKateSG — How Robust Statistics Work
- eduKateSG — How Sampling Distributions Work
Final Thought: Quantiles Tell Us Where Probability Crosses a Line
The mean asks where weighted magnitude balances.
The quantile asks where accumulated probability reaches a chosen level.
That simple change of question is powerful.
It can define a median.
A service promise.
A safety threshold.
A stock level.
A prediction bound.
A risk limit.
Quantiles are powerful because they turn probability into position. Their limitation is the same source as their robustness: once the position is known, much of the magnitude beyond it is allowed to disappear.