Part 2: Regimes, forecasts and tails · Chapter 6 of 14 · 8 min

Fat tails and extreme moves

Gold's hourly returns do not follow the bell curve: the kurtosis is 21.7 (the normal has 3). A move beyond three standard deviations happens in 2.0% of hours, against 0.27% for the normal (98 times against 13 expected); beyond four in 0.97%, against 0.006% (47 against 0.3). The 49 biggest candles (1%) make up 31% of the total variance. A candle with a true range above 3 ATR shows up one day in three.

The bell and its tails

Many finance calculations start from a convenient assumption: price movements follow a normal distribution, the bell curve. In this curve nearly everything sits near the centre and extreme values are very rare: beyond three standard deviations it happens 3 times in 1,000, beyond four once in 16,000.

Markets do not behave like that. The kurtosis measures how much the tails (the extreme values) weigh compared with the normal, which is 3. On gold's hourly returns it is 21.7. A fat-tailed distribution has two faces:

  • many moves smaller than normal (the centre is more peaked: 17.3% of hours sit beyond one standard deviation, against 31.7% for the normal);
  • a few enormous moves that a bell curve considers impossible.

How often, in numbers

The standard deviation of hourly returns is 0.369%. On the 4,864 hourly changes:

ThresholdHours beyond the thresholdShareNormalExpected from the normal
1 deviation84217.3%31.7%1,543
2 deviations2174.46%4.55%221
3 deviations982.01%0.27%13
4 deviations470.97%0.006%0.3
5 deviations220.45%almost zeroalmost zero

The meeting point is at two deviations, where the real data and the normal coincide (4.46% against 4.55%). From there on the normal gives up: at three deviations the data have 7.5 times more events, at four about 160 times (47 against 0.3). Twenty-two hours moved more than five deviations in just 211 market days, an event that for a normal would be seen once every three centuries.

On the 211 days (open to close) the phenomenon shows again, more muted: 12 days beyond two deviations (5.7%, against 4.55%) and 3 beyond three (1.4%, against 0.27%).

How much the extremes weigh

The 49 biggest candles (1% of the total) make up 31% of the total variance of hourly returns. The ten biggest, 13%. Measured on the true range instead, the ten biggest candles are worth 2.0% of the sum ($2,102 out of $107,543): $392, $204, $196, $196, $194, $190, $186, $183, $182 and $179. The two measures say different things: the variance weighs extremes squared and is dominated by them, the sum of the ranges counts them once. When a rule uses the square (standard deviation, Value at Risk), a handful of candles decides everything.

The biggest candle in the data is 29 January 2026 at 15:00 UTC: a true range of $391.5, almost 8 times the previous ATR ($49.7) and 23 times the median true range ($17).

In ATR: how «big» a candle is

Dividing each candle's true range by the ATR of the previous candle:

MeasureValue
Median0.87
90th percentile1.74
99th percentile3.65
Greater than 2 ATR319 candles (6.6%)
Greater than 3 ATR96 candles (2.0%)
Greater than 5 ATR22 candles

The median is below 1 because the ATR is an average and big candles raise it. The comparison with a series without clustering is instructive: shuffling the candles one by one (no clustering), true ranges above 3 ATR would be 186 on average instead of 96. In the real data the updated ATR partly absorbs the extremes: when the market gets agitated, the ATR rises and another big candle is no longer a surprise (chapter 3).

Figure 1. Each day's maximum true range divided by the ATR of the previous candle (bars): the line is at 3. On 73 days out of 211 at least one candle exceeds 3 ATR; the maximum of the period is 9.1 times.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, grey bars of the day's maximum of the ratio between true range and ATR and a horizontal line at 3.

On 73 days out of 211 (35%) at least one candle exceeded 3 ATR. A candle above 3 ATR therefore shows up about one day in three.

What it means for stops and risk

  1. A stop is not a ceiling on the loss. A candle like the late-January one (almost 8 ATR) can cut through the stop and fill it much further away: the gap between stop price and exit price is called slippage and it is worse in the tails. With a 1.5 ATR stop, an 8 ATR candle can cost even more than 5 R if the order is executed near the extreme.
  2. A bell-curve model understates extreme risk by an order of magnitude. If you take the standard deviation and derive «how much I lose in 99% of cases», in the tail the number is wrong on the low side.
  3. Rules based on the square are dominated by a few events: the standard deviation of a ten-month sample depends heavily on the ten biggest candles.
  4. Risk per trade must be chosen thinking of the worst candle, not the typical one: book 11 has already done this with size in R and drawdown.

What to do

  1. Measure the kurtosis and the shares beyond 3 and 4 deviations on your instrument's data: your normal is wrong in the tails.
  2. Look at true ranges in multiples of the ATR, not just in dollars: it says what counts as «big».
  3. Count the candles above 3 ATR (here 2%): it is how often the plan will get a surprise.
  4. Do not trust the standard deviation of a few months: a few candles decide it.
  5. Write in the plan what you do if the price jumps the stop: expected slippage, reduced size, no trades near news (book 9).

Procedure in four steps

  1. Calculate the standard deviation of hourly returns and count the hours beyond 2, 3, 4 and 5 deviations.
  2. Compare them with the normal (4.55%, 0.27%, 0.006%).
  3. Express true ranges in multiples of the previous ATR and count those above 3.
  4. Write in the plan what you do on the days when it happens.

Exercises

Chapter exercises

1. Hourly returns beyond three standard deviations are 2.01% in the data and 0.27% in a normal. How many times more frequent are they?

2. On about one day in three there is at least one candle with a true range above 3 ATR. What does it imply for a 1.5 ATR stop?

3. The ten biggest candles are worth 13% of the variance and 2.0% of the sum of true ranges. What does it say?

4. The kurtosis of hourly returns is 21.7 (the normal 3). How do you read it?

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.