Part 1: What volatility is · Chapter 3 of 14 · 8 min

Volatility clusters

On gold the direction of one hour says nothing about the direction of the next (autocorrelation 0.00), but its size does: between the true ranges of consecutive candles the autocorrelation is 0.59, it is still 0.38 at 24 hours and 0.12 at five days, against 0.00 for a shuffled series. A day above the median is followed by another above the median 62% of the time (after a day below, 38%). But almost all the daily persistence comes from January-March (0.69); from April it is 0.05.

Calm after calm, agitated after agitated

Whoever looks at a chart for long notices a fact: big moves come in bunches. After a violent hour another one arrives; after a quiet day a quiet one often follows. It is called volatility clustering and it is one of the most solid regularities of financial markets.

The clustering concerns the size, not the direction. An hour that rises a lot does not make a rising or a falling hour more likely; it makes a big hour more likely. That is why volatility can be estimated with some confidence while direction cannot.

How it is measured: autocorrelation

The autocorrelation at lag k measures how much the value of now resembles the one k steps ago: it is 0 if there is no link, 1 if the two values always move together. To know what 0 is worth «by chance», the candles are shuffled into random order (all links are removed) and the calculation is repeated 100 times.

On the 4,865 hourly candles:

LagReturn (direction)Absolute value of the returnShuffled series, 95th percentileTrue rangeShuffled series, 95th percentile
1 hour0.0000.2530.0390.5850.026
2 hours−0.0150.2070.0300.4910.025
5 hours0.0030.1580.0310.4120.025
24 hours−0.0030.1660.0260.3840.021
48 hours0.0050.0840.0320.2720.025
120 hours (5 days)−0.0200.0460.0290.1240.027

Three things.

  1. Direction has no memory: the autocorrelation of the return is zero at every lag (the wobbles around zero sit in the noise: the 95th percentile of the noise is about 0.03).
  2. Size has a lot: between the true ranges of consecutive candles it is 0.59 (against 0.00 for a shuffled series). For measures of this kind that is a very high correlation.
  3. It fades slowly: it falls below half of the one-hour value only at 48 hours, and at five days (0.12) it is still above the noise (0.03). An agitated hour says something about the volatility of the next two days, and a little about the next five.

On the day

With the daily ranges of the 211 market days:

LagAutocorrelation of the daily rangeOf the daily return
1 day0.629−0.034
2 days0.456−0.164
3 days0.313−0.002
5 days0.2840.065
10 days0.0060.089
20 days−0.030−0.035

With 211 days the error on the measurement is about ±0.13: the autocorrelation of the ranges stays well outside the noise up to five days and disappears at ten. The −0.164 of the return at two days leaves the band by a little, but of six lags tried one is expected every now and then, and there is no reason to read a rule into it.

Another way to say it: after a day with the range above the median, the next day is still above the median in 62% of cases (65 out of 105); after a day below the median, it is above in 38% (40 out of 105). With no link it would be 50% in both.

Figure 1. Daily range (grey bars) and 10-day mean (pink), in $. From mid-January to mid-April volatility stays high, then it falls and stays between $70 and $130 for the rest of the data.Data: gold, XAUUSDT perpetual future (real candles) · daily (D1) · from 15 Dec 2025 to 5 Oct 2026.
Read the chart in words

Daily gold chart with, below, the daily range bars and a moving-average line, with a grey band from January to April.

WhenWhatPrice
12 Jan 2026 → 15 Apr 2026January-April: volatility stays high

Times in UTC.

Careful: a single episode can dominate

The mean daily range per month is: December $78 (half a month), January $152, February $164, March $178, April $107, May $97, June $112, July $86, August $103, September $95, October $65 (six days). January-March is a single episode, with days of $464, $492 and $717.

If we split the one-day autocorrelation of the range:

PeriodOne-day autocorrelation
All data0.629
First 50 days (up to 20 February)0.762
January-March0.689
From 23 February on0.361
From April on0.047

From April, in a phase of ordinary volatility, the day-to-day persistence has practically disappeared. The hourly one remains, weaker: the autocorrelation of the hourly true ranges from April is 0.36 at one hour and 0.13 at 24 hours (before April 0.62 and 0.41). The picture changes: clustering is stronger when the market is already agitated, and a measure calculated on a period dominated by a single episode overstates the persistence of the rest.

This is a limit of our data, not a general rule: ten months are too few to know how often episodes like this happen.

What follows

  • For stop and size, the recent ATR is a good starting point: the volatility of now resembles that of yesterday and tomorrow.
  • Looking at the ATR of a month ago is not enough: the daily persistence disappears in about ten days.
  • There is no direction signal in volatility: a big candle does not say whether the next will rise or fall (chapter 12).
  • After an agitated period expect more agitation, for days; after a calm one, calm, but with less certainty.

What to do

  1. Update the volatility measure at every candle or every day, not once a month.
  2. Use the recent ATR (14 candles, or 10 days on the daily) for stop and size.
  3. Do not extrapolate beyond a week: after ten days the persistence on the daily range cannot be seen.
  4. Check the period before trusting a number: if one episode dominates, the measure is its.
  5. Do not use volatility to predict direction.

Procedure in four steps

  1. Calculate the autocorrelation of the true range (or of the daily range) at several lags.
  2. Compare it with a shuffled series, which gives the value due to chance.
  3. Split the data into periods and repeat the calculation.
  4. Use the persistence only over the horizon where the measure sits above chance.

Exercises

Chapter exercises

1. The autocorrelation of the hourly return is 0.00 and that of the true range 0.59. What does it mean?

2. After a day with the range above the median, the next day is above the median in 62% of cases, after a day below in 38%. What does it imply?

3. The one-day autocorrelation of the range is 0.69 in January-March and 0.05 from April. What lesson follows?

4. For how many days does the persistence of the daily range stay above the noise in our data?

Have a question or want to share your exercise?

Post in the community, or join the free signals room on Telegram.

Educational content, not financial advice. Trading involves risk.