Look at a chart that has climbed for six months and ask yourself whether it is trending. Most people answer that question by looking, and the answer seems obvious enough, because the line goes up and to the right.
The trouble is that the answer runs two different things together, one being where price ended up, and the other being whether the moves along the way had any memory of each other.
Every figure and number below is measured, not illustrative. The simulated markets are generated and read with the same measurement code that runs on the platform. The null and detection figures use 400 runs per scenario at 600 bars, the chance figure uses 1500 runs, and the nested-window comparison uses 500 runs per scenario. Where a figure states a reading, that reading came out of the measurement rather than being drawn to make a point.
Those come apart more often than you would expect. We generated a market with strong follow-through, where each move genuinely did tend to be followed by another in the same direction. Over 600 bars it finished down 58.3%, so the memory was real even though the direction went the other way, which is possible because the memory was never a statement about direction in the first place.
Three markets with the same volatility and very different memory
A name collision worth clearing up first
Two different things in market analysis carry the name Hurst, and confusing them is one of the most common ways to get lost early.
J.M. Hurst was an engineer who, in the 1970s, set out a theory of price behaviour built on cycles, working with envelopes, moving averages and the idea that a price move is the sum of several waves of different lengths. That body of work is what people mean by the cyclic tradition.
The Hurst exponent is a statistic. Harold Edwin Hurst was a hydrologist who spent decades measuring the Nile. In 1951 he published a way to measure whether a series has long memory, built from how the range of a series grows as you look at longer stretches of it. That is rescaled range analysis.
They are not the same person and not the same idea, and this article is about the second one.
One more word does double duty. Here, persistence means the memory between one move and the next. Elsewhere on this site it means whether a detected cycle keeps showing up in later data, which is a separate question with its own methods, covered in Cycle Persistence Testing.
What the number means
- Around 0.5, moves have no memory. What happened last is no guide to what happens next. This is what a coin flip looks like.
- Above 0.5, moves tend to follow through. A rise is a little more likely to be followed by another rise.
- Below 0.5, moves tend to snap back. A rise is a little more likely to be followed by a fall.
That is the whole idea, and it is narrower than it first sounds, because it says nothing about direction and nothing about tomorrow. What it describes is a tendency spread across many moves.
Mistake one: one number for a whole history
A single reading over ten years of data tells you about ten years, and markets do not hold one character for anything like that long.
If a market spent three years snapping back and two years trending, one number over the whole span reports something in between that was never true at any point. It is an average of two different worlds.
The fix sounds obvious: measure over shorter, recent stretches, which is what a rolling Hurst exponent does. But that opens a much harder problem, and it is the one the whole design turns on.
Mistake two: judging every stretch by the same yardstick
You will see the advice that a reading above 0.55 means trending and below 0.45 means mean reverting. It is the standard introduction to the subject, and it is where our own guides to trending versus ranging markets and market regime detection start a reader off. As a first orientation it is fine. As a rule applied to any amount of data it is wrong, and it gets worse the less data you have.
Measure 50 bars and the estimate comes out jumpy, because if you feed a market with genuinely no memory into the calculation over that short a stretch, the readings spread wide either side of 0.5 purely from chance. Measure 400 bars of the same memoryless market and those readings cluster much closer to 0.5. More data gives a steadier estimate, which is not a quirk of this method but simply how estimates behave everywhere.
How far from 0.5 does a reading have to be, at this amount of data, before chance stops being a good explanation for it?
You answer that by simulating, which means generating a very large number of markets known to have no memory at all, running the same measurement over them at each length, and seeing how far the readings actually spread.
What a market with no memory really produces
Read that top row again, because a fixed 0.55 rule would call a large share of pure noise a trend. Notice the shape as well: the range at 400 bars is about a third as wide as the range at 50, which is why one fixed band cannot be right at both ends.
What calibration actually buys
We ran 400 simulated memoryless markets and counted how often each length produced a false alarm. Then we ran 400 markets that did have a real but weak amount of follow-through, and counted how often each length caught it.
Flat false alarms, rising detection
That contrast is the argument for reading several lengths at once instead of picking one, since a short read reacts quickly but misses faint structure, while a long read sees faint structure but is slow to notice when things change. What you want is both of them, judged by the same standard.
Reading four lengths at once
So we measure four stretches: the last 50 bars, the last 100, the last 200, the last 400. One detail matters more than any other, and it is the thing readers most often get wrong when they first see it. These all end at the same bar, today.
Four nested windows, not four periods of history
What four readings can and cannot tell you
When several lengths agree, you have a market whose character holds across the amount of history you look at. That is worth more than any single reading, and it is what sits behind the regime wording described in Market Regime Detection.
If the shortest is outside the range and the longer ones are not, the behaviour is in the recent bars. That inference is fairly safe, because the longer reads have a narrower range and more power. If the effect ran through those extra bars too, they would usually have caught it.
The harder case runs the other way round. Suppose the 100 bar read is outside the range while the 50 bar read is not. It is tempting to conclude that the behaviour sits in bars 100 to 50 back, and that is one explanation, but a second one fits the same picture just as well. A longer read has a narrower range and more power, so what you are seeing may instead be a weaker effect spread across all 100 bars that 50 bars of data simply cannot resolve.
We tested how often each explanation produces exactly that picture, 500 runs each.
| What was really going on | How often it produced “100 outside, 50 inside” |
|---|---|
| Effect concentrated in bars 100 to 50 back | 18.0% |
| Weak effect spread across the whole series | 4.0% |
| Nothing at all, pure chance | 5.8% |
Read the third row against the first. The concentrated explanation is the most likely of the three, but plain chance is not far behind it. Anyone who tells you exactly where in time the behaviour sits, from four nested readings, is overselling. The honest statement is that the readings narrow it down and cannot settle it.
The mistake almost nobody mentions
This one matters most, and it applies to every method of this kind rather than only to ours. We check four lengths, and each of them has roughly a one in ten chance of a false alarm on a market with no memory, so what is the chance that at least one of the four fires by chance alone?
What chance alone produces across four checks
That number is worth sitting with, because it means that if you check four things and act on whichever one lights up, markets with no memory at all will fool you roughly one time in four.
This is not a flaw in the measurement, but simply what happens whenever you run several tests and pay attention to whichever one passes. It is also the reason a single length outside the range is treated as the weakest kind of evidence rather than a settled result, and why the wording only escalates as more lengths agree. Most published work on this topic skips the step quietly, whereas we put the number on the screen next to the reading.
Drift is not persistence
This brings us back to where we started, because it is the most common objection we get. Someone looks at a chart that has climbed steadily, sees a reading near 0.5, and concludes that the measurement must be broken. It is not broken, and the reason is that the chart and the reading are answering two different questions.
A market with no memory at all that still climbed 192.7%
Figure 1 makes the same point from the other side, since the series with the strongest follow-through finished down 58.3%, which is strong memory and a negative return in the same series. Put the two figures together and the lesson is that a rising chart is not evidence of follow-through, and follow-through is not a promise of a rising chart. So if a chart slopes while the persistence reading stays quiet, that is information rather than an error.
How to read four numbers
Everything above establishes what a reading is worth. The sequence that keeps the four straight, and keeps the multiplicity problem in view, is short.
Ask what the length can resolve before you read the number
A 50 bar reading and a 400 bar reading are not the same instrument, because the short one reacts quickly and spreads wide, whereas the long one is narrow and slow. The number only means something once you know which of the two produced it.
Count how many lengths agree
Agreement across lengths is the strongest thing this measurement offers, because it says the character holds across the amount of history you look at. One length alone is the weakest, and roughly one memoryless market in four produces exactly that.
Prefer the safe inference over the tempting one
A short length outside the range with the longer ones inside points at the recent bars, and that inference holds up. The reverse case, a longer length outside with the shorter one inside, has more than one explanation and the readings cannot settle which.
Read the result as character, not as direction
Follow-through is not a forecast of a rising chart, and a rising chart is not evidence of follow-through. What the reading does is change how much weight your other analysis deserves, rather than replace that analysis.
What this is not
Nothing described here is a buy or sell signal, and nothing on the FractalCycles platform is. This measures the character of a market, not what it will do next. It does not forecast direction, it does not time turns, and a market can change character the day after you measure it.
What it is good for is context, because knowing whether the moves in front of you have been following through or snapping back, and knowing how much evidence stands behind that, changes how much weight any other analysis deserves.
Common questions
- What does the Hurst exponent actually measure? Whether moves in a series tend to be followed by more moves in the same direction. Around 0.5 there is no memory. Above 0.5 moves tend to follow through. Below 0.5 they tend to snap back. It says nothing about direction and nothing about what happens next.
- Is a Hurst reading above 0.55 a trend? Not on its own. Over 50 bars, a market with no memory at all routinely produces readings between 0.335 and 0.665 purely by chance. A fixed 0.55 rule labels a large share of pure noise as trending. The threshold has to depend on how many bars were measured.
- Why does the threshold change with the number of bars? More data makes the estimate steadier. The range a memoryless market stays inside is 0.331 wide at 50 bars and 0.118 wide at 400 bars, about a third as wide. A single fixed band cannot be correct at both ends.
- Can a market trend while the persistence reading is quiet? Yes, because drift and persistence are different things, and price can climb steadily while each individual move is unrelated to the last. In our own test the series with the strongest follow-through finished down 58.3%, which is memory and a falling price in the same series.
- Is J.M. Hurst the same as the Hurst exponent? No. J.M. Hurst wrote about market cycles in the 1970s, whereas the Hurst exponent comes from Harold Edwin Hurst, a hydrologist who published rescaled range analysis in 1951 after studying Nile flow records. They are different people with different ideas.
Sources
- Hurst, H.E. (1951). Long term storage capacity of reservoirs. The original rescaled range work, from Nile flow records.
- Bartels, R. (1932). On the significance testing of periodic phenomena. The tradition of asking whether an apparent pattern survives a chance explanation.
- Hurst, J.M. (1970s). The cyclic principles behind the separate cycle tradition, distinct from the exponent above.
- Simulation figures were generated with the production measurement code. Null and power figures use 400 runs per scenario at 600 bars. The chance figure uses 1500 runs. The nested-window comparison uses 500 runs per scenario.