What a Percentile Hides
A rank, not a distance
A percentile reports the share of a sample that falls at or below a value. If a fund's return sits in the 75th percentile, then 75% of the peer group did worse and 25% did better. That is the entire content of the statement.
Notice what it does not say. It does not say how far the value is from the middle, how likely a repeat is, or whether the underlying thing has changed. It is a position, and position is only meaningful relative to a defined population.
This is why a percentile is simultaneously the most useful summary in macro and the easiest one to abuse. It compresses a whole distribution into one digit — and it silently carries the choice of that distribution with it.
Worked example: our own curve
Take the 2s10s spread against the 2026 record. From 2 January to 21 September there are 181 observations on Treasury's par yield curve. The spread's low for that period is 20.0 basis points, its median is 50.0, and its high is 74.0.
On 21 September the spread closed at 20.0 basis points. That is not near the bottom of the year's range — it is the bottom. It is also the only session in 2026 that printed below 25 basis points.
Now hold that against the same series 3 weeks earlier. On 4 September the spread was 41.0 basis points, which put it at the 24th percentile of the year. So a move from 41 to 20 — 21 basis points of price — carried the spread from the lower quarter of its distribution to the floor of it.
That is the value of a percentile: it translates price into position. A 21 basis point move sounds routine. Moving from the 24th percentile to the minimum of 181 observations is not routine, and only the percentile tells you so.
The window is the whole argument
The same 20 basis points that is the lowest close of 2026 is unremarkable against a decade, and would be flatly inverted against the 1970s. Nothing about the number changed. The window did.
This site has a rule about it, written into the 17 September rates piece: "Anyone claiming 'the long-run average' without naming the window is choosing a number that suits the conclusion." The same discipline has to apply to percentiles, because a percentile is that rule with a statistic attached.
Here is what the failure looks like in practice — from our own archive. Three pieces published within 5 days described the 30yr yield as sitting at the 99th, the 98.6th and the 99.3rd percentile. None of them was lying. Each used a different window, and not one printed the window. For a reader, three numbers for one fact is indistinguishable from three errors.
So the house standard is: a percentile does not leave the building without its series, its window and its sample size. Three components, every time. If any is missing, you are reading an opinion in a statistic's clothing.
Where percentiles lie
Small samples. A tenth percentile needs enough observations to have a tenth. Claims at the 4th percentile of a series running back to 1967 are meaningful; the same claim against 3 years of stored data is a much weaker sentence, even when the number looks identical.
The 100th percentile does not exist. If a value is the maximum of a finite sample, the share of observations strictly below it is one less than the total — never the whole set. The 2yr yield closed at 4.76% on 21 September, which is the highest print of 2026: it is the maximum, and describing it as the 100th percentile is arithmetically sloppy. That sloppiness is precisely how one fact becomes three numbers.
A percentile assumes the past is the population. It describes a distribution that has already happened. When the regime changes, the distribution changes with it, and a reading that was a warning can quietly become the new normal. The 2yr at its 2026 high is not stretched relative to a long-run mean — it is simply the highest value observed. "High" and "about to revert" are different claims, and the percentile supports only the first.
How we use them
Percentiles do real work on this site, always with the window attached. High-yield credit spreads at 2.70%, the 10th percentile of the roughly 3 years this site stores, against a record tight of 2.59% in that same window. Initial jobless claims at 206k, the 4th percentile of a series that runs back to 1967. The 10yr real yield at 2.62%, a level reached or exceeded in 69 of the 5,930 daily observations in the TIPS sample — 1.2% of it.
Look at what the third one does that the others cannot: it states the count and the sample. Anyone can check it. That is the standard — not because it is more precise, but because it is falsifiable by a reader who wants to argue.