Part 1: Risk and its numbers · Chapter 2 of 14 · 12 min

The series of results

A series of trades is a sample: run again, it would give a different sequence. With a deviation of 1 R the error on the mean is ±0.19 R after 100 trades and ±0.08 after 552. In our series, 100 consecutive trades went from −24 to +11 R (levels bounce) and from −16 to +19 R (swing break); even the example with a true edge of 0.21 R had a block of 112 trades at +8 R against another at +36.

Your series is one of many possible orders

You made 100 trades and the account is up 8 R. Do you have a system that works? To answer you need to know how often a system with no edge at all produces +8 R in 100 trades. If it happens one time in three, your +8 R says nothing; if it happens one time in a thousand, it says something.

The results of a system are like hands in a card game: each trade depends on that day's market, and the same system, applied over the same ten months in a different order or on slightly different days, would give a different sequence. The series in front of you is a sample of all the series the system could have produced. This chapter measures how far sample and reality can diverge.

Mean, deviation, error

Three numbers describe a series of results in R:

  • the mean (mean R per trade): the sum of results divided by the number of trades;
  • the standard deviation: how far a typical result sits from the mean;
  • the error, that is the uncertainty on the mean: ±1.96 times the deviation divided by the square root of the number of trades. It tells you where the true mean would fall in 95% of cases, if the series were a random sample of the true one.

For series LA the deviation is 0.99 R. The size of the error depends on how many trades you have:

Trades1025501002005001,000
Error on LA (sd 0.99)±0.62±0.39±0.28±0.19±0.14±0.09±0.06
Error on LD (sd 1.00)±0.62±0.39±0.28±0.20±0.14±0.09±0.06
Error on SD (sd 0.56)±0.35±0.22±0.15±0.11±0.08±0.05±0.03

The error falls with the square root: to halve it you need four times as many trades. After 25 trades at a deviation of 1 R, a mean R of +0.2 has an error of ±0.39: it could be +0.6 as well as −0.2. In a system like this you need about 100 trades for a +0.2 R mean to be just distinguishable from zero (error ±0.2), and about 400 to tell it apart with a margin (error ±0.10). Chapter 12 derives the rule precisely.

One hundred trades are few

Take the real series and slide a window of 100 consecutive trades along, one trade at a time, looking at the sum of R in each window:

SeriesWindowsMinimum sumMaximum sumMedianWindows in profitAbove +10 RBelow −10 R
LA (mean −0.03)453−24.1+10.8−0.842%2%18%
LD (mean +0.03)268−15.8+19.2+6.375%39%5%
SX (mean +0.21, true edge)453+6.0+38.0+22.0100%94%0%

Three readings, in order:

  1. LA has no edge and had windows at +10.8 R. A person who had started at the right moment and stopped at the peak would have sworn to have a system.
  2. LD has an almost null mean (+0.03) and yet 75% of its windows are in profit. The windows overlap, they are not 268 independent trials: they are about three different stretches of the series, one of them good. Counting windows in profit as if they were independent results would be a mistake.
  3. SX has a true edge and all windows are positive, but the sum ranges from +6 to +38 R: a gap of 32 R between the best and the worst hundred trades of the same system.
Figure 1. The sum of R of the last 100 trades day by day, from when 100 exist (December 2025 - October 2026): levels bounce (bars), swing break (pink), synthetic example with an edge (blue). The real series cross above and below zero; the synthetic one stays above zero but swings.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, three curves of the rolling sum of R over 100 trades.

Consecutive blocks, same system

Split each series into consecutive blocks with the same number of trades and add up R in each:

SeriesBlock 1Block 2Block 3Block 4Block 5
LA (110 trades per block)+12.2 R−0.3 R−1.8 R−6.9 R−21.4 R
LD (91-94 trades per block, 4 blocks)+2.2 R+13.7 R+4.3 R−9.5 R
SX (110 trades per block)+24.0 R+26.0 R+20.0 R+36.0 R+8.0 R

LA goes from +12.2 R in the first block to −21.4 R in the last, an almost continuous slide. A careful reader asks whether the system "stopped working". With a deviation of 1 R and 110 trades per block, a block has a deviation of about 10.4 R; with a true mean of −0.03 you expect −3 R per block, so the first block sits about 1.5 deviations above, the last about 1.7 below. It is a large swing but compatible with five blocks drawn from the same distribution, and with five blocks nothing more can be said.

The same holds for SX, which has a built-in edge that we know: block 5 at +8 R (mean +0.07 per trade) against block 4 at +36 R (mean +0.33). Had we not known the edge was true, we would have spoken of a "system that wore out".

Figure 2. Cumulative R at the same instants (December 2025 - October 2026): levels bounce (bars, −18.2 R final) and the synthetic example SX with an edge (blue, +114 R final). Even the curve with the true edge has flat stretches: the last block of 112 trades earns only +8 R.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, two curves of cumulative R.

Drawing at random: the bootstrap

One way to measure how much a series depends on the order is the bootstrap. You take all the observed results, draw at random with replacement (each result can come out more than once) as many results as you need, and add them up. You repeat 20,000 times. You get the distribution of sums that the same series would give in different orders and combinations.

For 100 trades, the result is:

SeriesPositive sums5th percentileMedian95th percentileAbove +10 RBelow −10 R
LA37%−19.6−3.5+12.99%25%
LD62%−13.5+3.1+19.324%9%
SD95%+0.1+9.4+18.446%0%
SX98%+4.0+20.0+36.084%0%

(The 5th and 95th percentiles enclose 90% of the sums.) For LA, which has a mean of −0.03: a hundred trades give a positive result one time in three (37%), and +10 R or more one time in eleven (9%). Whoever looks at the first 100 results and sees +10 R has no proof of anything.

With 552 trades, the whole series, the bootstrap for LA still gives 21% of positive sums (from −56 to +20 R between the 5th and 95th percentile); for SX 100% (from +76 to +154).

The bootstrap has a limit that must be stated: it draws from what you observed. If the observed series was lucky, the bootstrap treats it as the truth. For SD, with only 55 trades, the 95% of positive sums does not hold: the error on the mean (±0.15) includes zero, and the bootstrap ignores it. When trades are few, the only remedy is to collect more.

Mean and median

In a series of ±1 R the median is of no use. LA has a median of −0.81 R, because more than half of the trades are a loss close to −1; but the mean is −0.03, because the gains close to +1 are almost as numerous. With results concentrated at ±1, the median takes the value of the coin that falls on the more frequent side and says nothing about the sum. For a system with small losses and rare, large gains the opposite happens. In both cases you need the mean, with its error.

What to do with it

  1. Compute mean, deviation and error after every 25 trades, not every day. The error tells you how far you can trust the mean.
  2. Do not look at small blocks. Ten or twenty trades have an error of ±0.4-0.6 R, larger than any realistic edge.
  3. Compare with the bootstrap: if more than one random sum in ten does better than your result, your result is not enough to convince.
  4. Count independent stretches, not rolling windows. A hundred overlapping windows are still one series.
  5. Do not infer a change of regime from a single block. You need several blocks, and a comparison with the swing you expect.

Procedure in four steps

  1. Record every result in R, in the order it happens.
  2. Every 25 trades compute mean, deviation and error.
  3. Compare the sum of the last 100 trades with the distribution obtained from the bootstrap on your series.
  4. Decide in advance the number of trades beyond which you will change something, and do not change it halfway.

Exercises

Chapter exercises

1. A system has a deviation of 1 R. After 25 trades its mean R is +0.2. What can you say?

2. To halve the error on the mean, how many trades do you need compared with before?

3. SX has a true edge of about 0.21 R per trade. In the last block of 112 trades it made +8 R (mean 0.07). What does it mean?

4. With LA (mean −0.03) the bootstrap gives 37% positive sums over 100 trades. After 100 trades your system made +8 R. What do you infer?

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.