Percentile

If you've ever wanted to say "today is unusually high" and actually *mean* something measurable by it, a percentile is one of the simplest ways to do that.

In 30 Seconds

• Ranks a value against its history — the 90th percentile means higher than 90% of recent readings.
• Easy to compare across markets — a 95th-percentile reading means the same thing whether it's Gold, NASDAQ, or ES, despite very different raw numbers.
• Insensitive to outliers — one freak day barely moves a percentile, unlike an average.
• Does not predict direction — it describes how unusual today's reading is, not the next move.
• HIE discloses exactly how it's computed — the window and method behind every percentile.

Definition

A percentile tells you where a value ranks within a set of past values, expressed as the percentage of readings that fall below it. If today's volume is at the 80th percentile of the last 250 sessions, it means today's volume was higher than 80% of those sessions — and lower than the top 20%.

"The idea is rank, not distance."

Unlike a z-score (which measures *how far* a value is from average in standard-deviation units), a percentile measures *where a value sits in the ordering* of its history. That difference matters: percentiles don't assume any particular distribution shape, which makes them robust when data is skewed or has fat tails — exactly the conditions markets tend to produce.

Two things define any percentile: the set of past values it's ranked against (which window of history), and the method used to compute the rank. As with any relative statistic, a percentile without its window disclosed cannot be interpreted correctly — "80th percentile of what?" is always the first question.

Why It Matters

Percentiles are often the most honest way to say whether something is "high" or "low," because they sidestep the two biggest weaknesses of averages.

First, percentiles resist outliers. Imagine one unusually large volume spike. An average can move noticeably because of that single event — enough to make normal days look quiet by comparison. A percentile usually changes very little, because one extreme observation affects only one position in the ranking; the rest of the distribution keeps its shape. When you want a stable sense of "is this unusual," rank beats average.

Second, percentiles don't assume a bell curve. A z-score's clean interpretation leans on the normal distribution, which markets routinely violate. A percentile makes no such assumption — it simply asks "how many past readings were below this one?" — so it stays meaningful on skewed, lumpy, fat-tailed market data.

Traders often use percentiles to gauge unusually high or low trading volume, unusually wide or narrow ranges, how stretched price is from a reference, and other conditions where "compared to normal" is the real question. To ground it with no math: if you want to study "what tends to happen after an exceptionally quiet volume day," you first need an honest, stable definition of "exceptionally quiet." "Below the 5th percentile of the last 250 sessions" is that definition — and it stays meaningful whether the market is calm or wild, because it's always relative to the market's own recent behaviour.

How HIE Computes It

Two platforms can report different percentiles for the same reading, because they may rank against different windows or use different methods. That's why HIE discloses its methodology rather than presenting a bare number.

Percentile is a live, askable input in HIE, computed with a single disclosed, deterministic method — and every answer prints the exact window it used, so nothing rests on the reader knowing the defaults. Two properties are worth stating plainly:

• Rank within a strictly-trailing window. HIE ranks a value only against its own trailing history — the calculation at any bar uses nothing from after that bar. That is what "no look-ahead" means: a reading is ranked only against history that genuinely preceded it.
• A disclosed default window per feature, which you can override. Each feature carries its own default window from HIE's registry, and you can always specify your own (for instance, "over the last 250 bars"). Because the window depends on the feature, there's no single universal number — every disclosure states the actual window used.

HIE doesn't just name the method — it discloses the actual formula. In HIE's own disclosure ("what ran"), a percentile computation appears in this exact form:

PERCENTILE(VOLUME, window=250, method=rank within trailing window: 100*(#window <= current)/window, rolling, strictly-trailing, v1) < 5

Read plainly, that method line says: the percentile is the percentage of values in the trailing window that are *at or below* the current value. Nothing is hidden — the window (250), the exact ranking formula, the fact that it's rolling and strictly-trailing, and the version (v1) are all printed. So a question about volume below the 5th percentile over 250 bars resolves to, and is disclosed as, exactly the line above. The words you ask are shown as the precise computation they became. (Phrasings like "percentile," "percent rank," "top/bottom X%," and "in the lowest 5%" resolve to this same computation.)

Interpreting a Percentile

A percentile reads directly as "higher than this share of recent readings":

• 50th percentile — the median; exactly typical, half of recent readings were higher, half lower.
• 80th–95th — elevated to high; only a minority of recent readings exceeded it.
• Above 95th — unusually high relative to recent history.
• Below 5th — unusually low relative to recent history.

These thresholds are conventions, not universal rules — different studies may choose different cut-offs depending on the question being investigated.

Because a percentile is pure rank, it carries no assumption about *how far apart* the values are — only their order. That's its strength (robustness) and its limit (it won't tell you the *magnitude* of the gap, only the rank). For "how far from normal in size," a z-score is the better tool; for "how unusual in rank, robustly," a percentile is.

Common Mistakes

"It's in the 95th percentile, so it must come back down."

A percentile measures how unusual a current reading is — not what happens next. A 95th-percentile reading describes the present, not a signal that a reversal is coming. Treating "unusual" as an automatic reversal cue is the most common error.

"I changed the window and the percentile changed — something's wrong."

Expected, not wrong. A percentile ranks against whatever trailing window is used. A longer window includes more varied history, so the same value can rank differently. The window is part of the definition.

"90th percentile means a 90% chance of something."

No. A percentile is a rank, not a probability of a future event. "90th percentile" means higher than 90% of past readings — it says nothing about the odds of anything happening next.

"This is the 80th percentile here but the 60th on another site — one is wrong."

Not necessarily. Two percentiles computed over different windows or methods aren't comparable. A percentile is only meaningful alongside the window and method that produced it.

"A 95th-percentile reading is extremely rare."

Not in the way people mean. With a rolling window, the market naturally produces new high-percentile readings over time — a 95th-percentile day isn't a once-in-a-lifetime event, it's simply higher than 95% of the *recent* window, and windows move forward. High percentile means "unusual lately," not "unrepeatable."

Limitations

A percentile is only as meaningful as the window behind it. Over a short trailing window, the rank is based on few readings, so it can jump around on little evidence — the same small-sample caution that applies to every statistic applies here. (See Sample Size.)

A percentile deliberately focuses on rank rather than magnitude. It tells you a reading ranks near the top, but not by how much — two very different values can share a percentile if the ones between them are sparse. And like every relative measure, it describes the past window it was computed over; it is not a forecast. A reading in the 99th percentile was unusual relative to recent history — that is a description, not a prediction of what follows.

Try These HIE Questions

The best way to understand percentile is to test it against history. Because percentile is a live, askable input in HIE, this isn't just theory — you can put a historical question to it directly. Each of these runs in HIE today and returns a full certified answer, disclosing the exact computation it used:

• "ES volume below the 5th percentile" — resolves to `PERCENTILE(VOLUME, window=250, …) < 5`.
• "ES volume above the 95th percentile" — resolves to `PERCENTILE(VOLUME, window=250, …) > 95`.
• "ES volume top 5 percent" — resolves to the same `> 95` computation as the previous question: two different phrasings, one identical computation.
• "NQ price above the 95th percentile" — with no window specified, HIE applied price's default of 100 bars and disclosed it: `PERCENTILE(PRICE, window=100, …) > 95`. (A good illustration of the default-plus-disclosure behaviour: you didn't choose 100 — the registry did, and told you.)
• "NQ ATR below the 5th percentile" — resolves to `PERCENTILE(ATR, window=20, …) < 5`.
• "NQ RSI above the 95th percentile" (250-bar window chosen) — resolves to `PERCENTILE(RSI, window=250, …) > 95`.

*(These are real answers generated on 16–17 July 2026, research through those dates; because the historical data is periodically refreshed, the exact counts can change over time.)*

Percentile applies equally to price, volume, RSI, and ATR — the same computation on whichever feature you ask about, as the examples above show across all four.

One phrasing behaviour worth knowing: a question like "in its 95th percentile" deliberately does *not* run, because it's ambiguous — above the 95th, or below it? Rather than guess, HIE asks, with both options one tap away.

And a deliberate refusal worth understanding: if you ask HIE for a "rank" (for example, "NQ rank above 90"), it won't quietly treat that as a percentile. Rank (an ordinal position) and percentile (a 0–100 scale) are related but *not* the same thing, and HIE refuses to silently swap one for the other. Instead it asks whether you meant the percentile — one tap into the percentile flow — and you're not charged for the question. That refusal is the honesty principle in miniature: rather than answer a subtly different question, HIE names the difference and lets you choose.

How This Fits Into MPM

Percentile is one of the statistical building blocks behind HIE's ability to condition historical research on how unusual a reading was — turning vague ideas like "a quiet volume day" into precise, disclosed, testable definitions.

You'll encounter it in:

• HIE (Historical Intelligence Engine), where percentile is a live, askable input: you can ask historical questions conditioned on where a reading ranked, and every answer discloses exactly how the percentile was computed — the window and method — so the definition behind your question is never hidden. Percentile is live across every entry point (chat, the Reports page, the full report, matching charts, and the exported PDF and CSV), all running the same shared parser, so identical words produce an identical computation every time.
• Research Papers, where "unusually high/low" is expressed in disclosed percentile (or z-score) terms rather than raw points, so a claim holds up across calm and volatile eras.
• Intelligence Circle — a member-only research environment containing advanced market research, historical investigations, and trading strategies developed using the MPM research framework.

MPM's broader stance is that a measure of "unusual" is only honest if the yardstick is disclosed — which is why HIE states the window and method behind every percentile rather than presenting an unexplained number.