moonboard.ioResearch & Data
Learn

What is percentile ranking?

The single most important idea for reading Moonboard numbers correctly - and the one most often misread.

The definition

A percentile rank reports what share of a comparison set falls below a given value. Nothing else.

percentile = count(values below x) / count(all values) × 100

If 340 of 400 tracked assets have lower turnover than a given asset, that asset's turnover percentile is 85. The number describes its position in a queue, not the size of anything it holds.

Why it is used

Because the underlying quantities are unusable in raw form. Crypto data is heavy-tailed: the largest asset might have a thousand times the interactions of the median, and one asset can shift a mean by more than the other several hundred combined.

Percentiles solve this in one step. They are immune to outliers — the largest value gets rank 100 whether it is twice the runner-up or ten thousand times it. They require no distributional assumption. And they put every metric on a shared 0–100 scale, which makes different quantities comparable at a glance.

What it costs: magnitude

The strength is also the weakness. Percentiles preserve order and discard distance.

Universe A — clustered values: 10, 11, 12, 13, 14, 15, 16, 17, 18, 100 the 100 ranks at 90 — and is more than five times the next value Universe B — spread values: 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 the 100 also ranks at 90 — and is barely above its neighbour

Same percentile, entirely different situations. In one the leader is a dramatic outlier; in the other it is marginally ahead. A percentile cannot tell you which case you are in.

Z-scores take the opposite trade: they keep magnitude and pay for it with sensitivity to outliers. COR is the one Moonboard metric that makes that choice.

The property that surprises people

A percentile can change when the asset does not.

The rank is computed against the current comparison set. If other assets' values fall, an unchanged asset moves up. If the set gains new members above it, it moves down. Both happen without any new information about the asset itself.

Monday universe: 400 assets, our asset ranks 320th → percentile 80 Tuesday 40 assets see a burst of activity and overtake it our asset's own data is identical → percentile 70 The 10-point drop describes the other 40 assets.

This is not a flaw in the implementation; it is what a relative measure means. But it does mean a percentile time series mixes two signals — the asset's trajectory and the universe's — and separating them requires the raw values, not the ranks.

Where Moonboard uses it

MetricScaleRelative to the universe?
Moonboard ScoreAbsoluteNo — fixed thresholds throughout
OPSPercentileYes
LMRPercentileYes
SMIPercentileYes
SDEPercentileYes
CORCentred z-scoreYes — both position and scale
WSDCentredYes — both legs are percentiles

Six of the seven depend on the universe. Only the Moonboard Score can be compared across time and across different asset sets without qualification.

Reading percentile metrics well

  • Never read a percentile as a quantity. An LMR of 90 says nine in ten assets trade less actively. It says nothing about how much trades.
  • Ask what the universe is. "Top decile" across 400 mixed assets is a different claim from "top decile" across 40 large caps.
  • Treat small movements as noise. A shift from 71 to 74 is usually the field rearranging, not the asset changing.
  • Check the raw fields for anything important. The rank is a summary; the underlying values are the evidence.
Why this makes historical analysis hard

Reproducing a percentile from a past date requires reconstructing the entire universe as it stood then — every asset that was tracked, with the values it had, including assets that have since been delisted. That is a substantially harder problem than storing one asset's history, and it is the main reason no backtests are published here. See research notes.

Related