Part 2: Regimes, forecasts and tails · Chapter 5 of 14 · 8 min

Forecasting tomorrow's range

Six ways of estimating tomorrow's range (yesterday, 5, 14 and 50-day averages, an exponential average, the median of the whole past) give almost the same error: over the 161 test days the 14-day average is off by a typical 27% and the real range sits between half and one and a half times the estimate 83% of the time, against 81% for the simple median of the past. The correlation between estimate and reality is 0.33. For the range you think in intervals, not in single numbers.

What is forecast and what is not

In chapter 3 we saw that volatility has memory, in chapter 4 that regimes last days. A practical question follows: how much will gold move tomorrow? What is forecast is the size of the movement, not the direction. To set a stop, choose a realistic target or decide the size, you need a number for the day's range.

We tried six methods, all calculated only with the days before the one to forecast:

MethodHow it is calculated
Yesterday's rangethe previous day's range
5-day averagemean of the ranges of the last 5 days
14-day averagemean of the last 14 days
50-day averagemean of the last 50 days
Exponential averageeach day weighs 94% of the more recent one, and so on backwards
Median of the whole pastthe central value of all the previous ranges (no forecasting: just «the typical day»)

The tests start from the fifty-first day, to give the long methods the data they need: 161 test days remain, from 23 February to 5 October. So the most agitated part (January-February) is used to build the estimates but does not enter the tests.

How much they err

For each day the estimate is compared with the range that then occurred. The typical error is the median of |log(estimate ÷ real)|: a 27% means that the estimate is usually off by a little more than a quarter, up or down.

MethodCorrelation with the realTypical errorEstimate ÷ real (on average)Real between ½ and 1½ of the estimateReal below halfReal above 1½
Yesterday's range0.36335%1.0171%7%22%
5-day average0.34429%1.0882%5%13%
14-day average0.33427%1.1183%6%11%
50-day average0.34929%1.2182%9%9%
Exponential average (0.94)0.34731%1.1783%9%8%
Median of the whole past0.08428%1.0381%4%15%

Four readings.

  1. The 14-day average is the best, but only just: typical error 27% against 28% for the median of the past, share inside the interval 83% against 81%. The extra work on the other five pays nothing measurable in this data.
  2. The correlation is low: 0.33 means that the estimate explains about 11% (0.33 squared) of the differences in range from one day to the next. The rest is news.
  3. Yesterday's range is the worst: it uses a single datum, and one day's range is very noisy.
  4. All the averages tend to overestimate (estimate ÷ real above 1): the average is pulled up by the very agitated days, while the typical day is calmer. With 50 days the overestimate is 21%.
Figure 1. The day's range (grey bars) and estimate = mean of the 14 previous days (pink line), in $: the estimate follows the general level but not the single days, and it tends to sit above most of the bars.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 bars of each day's range and a smoother pink line.

An interval instead of a number

The ratio between the real range and the 14-day estimate, over the 161 days:

PercentileReal ÷ estimate
10th0.56
25th0.68
Median0.89
75th1.17
90th1.51

So, with the 14-day estimate, the day's range is:

  • between 0.56 and 1.51 times the estimate, eight times out of ten (10th to 90th percentile);
  • 0.89 times the estimate on median, that is about a tenth less;
  • above one and a half times the estimate on 11% of days, below half on 6%: the errors on the upside are almost double those on the downside.

For a stop, the last line matters more than the others: one time in nine the day is much bigger than expected, and a stop sized on the estimate is swept away.

What follows

  • For tomorrow's range there is no sharp number: there is a band (from half to one and a half times the 14-day average, 83% of the time).
  • A simple method is enough: the 14-day average or even the median of the past; complicating does not help in this data.
  • The uncertainty is asymmetric: days much more agitated than expected are double the days much calmer.
  • The forecast changes the scale, not the direction: it does not say whether to go long or short.

What to do

  1. Write an estimate of the day's range (for example the mean of the 14 previous days) and the band (0.5-1.5 times).
  2. Do not trust a single day: yesterday's range errs the most.
  3. Size stops and targets with the band, not with the point estimate.
  4. Put in the plan what you do on the 11% of days when the range is above one and a half times: a wide stop, less size or no trades.
  5. Retest every quarter: with more data the band changes.

Procedure in four steps

  1. Calculate the mean of the ranges of the last 14 days (closed days only).
  2. Multiply it by 0.5 and by 1.5 to get the band.
  3. Check every month how often the real range is inside the band (here 83%).
  4. Write in the plan the rule for days outside the band.

Exercises

Chapter exercises

1. The mean of the previous 14 days gives $100. In which band does the day's range sit 83% of the time?

2. The 14-day average errs by a typical 27%, the median of the whole past by 28%. What do you conclude?

3. The correlation between estimate and real range is 0.33. What share of the differences in range from one day to the next does it explain?

4. Why do errors on the upside (range above one and a half times the estimate, 11%) matter more for a stop than those on the downside (below half, 6%)?

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.