Part 4: Strategies, system and reference · Chapter 13 of 14 · 13 min
A time-based system on trial
We built a system with the rules of the book: breakout of the Asian range, only if the range is at least the median of the previous twenty days, skipping first Fridays and Fed decisions. Over 55 trades it gives +0.09 ± 0.15 R (random direction 0.00) and the best of the four variants tried is exceeded by 30% of random series. It is not proven: with this error 477 trades would be needed for ± 0.05, that is seven years.
From rules to a system
In chapters 2-10 we collected facts about times; in chapter 11 we saw that no entry rule beats chance. A system is the coherent whole: an entry, an exit, a risk management and filters chosen before looking at the results. Ours starts from the simplest rule of chapter 11, the breakout of the Asian range, and adds two filters that come straight from the previous chapters.
- Calendar filter (chapters 7-9). Skip the first Fridays of the month (the employment report day, usually) and Fed decision days. On those days the data or Fed candle has a size that no normal stop can withstand.
- Energy filter (chapter 3). The Asian range is the day's thermometer: enter only if the range is at least the median of the previous twenty days, that is if the day starts "hot".
The complete rules:
| Element | Rule |
|---|---|
| Range | High and low of the H1 candles between 00:00 and 07:00 UTC |
| Entry | At the close of the first H1 candle that closes beyond an extreme, between 07:00 and 15:59 UTC |
| Stop | At the opposite extreme of the range |
| Target | 1 R |
| Time exit | 21:00 UTC |
| Frequency | One trade a day, one position at a time (it always closes by the evening) |
| Calendar filter | Nothing on the first Friday of the month, nothing on the Fed decision day |
| Energy filter | Only if the Asian range is at least the median of the previous 20 days |
| Costs | 0.3 $ per trade |
How it is tested
The system has a methodological flaw that we declare at once: we looked at four variants, not one, because we want to show you the effect of each filter. So the test must take into account that we are choosing the best of four.
- A: the base rule, every day.
- B: A, without first Fridays and without Fed decisions.
- C: A, only when the Asian range is at least the median of the previous twenty days.
- D: B and C together: the system.
For each: the results, the two halves of the sample, random direction on the same entries and 300 random series (the same candles shuffled in blocks of six hours) on which we redid everything, filters included.
What happened
| Variant | Trades | Mean R ± error | Winners | Random direction | First half | Second half |
|---|---|---|---|---|---|---|
| A: every day | 169 | +0.05 ± 0.10 | 51% | −0.01 | +0.05 ± 0.15 | +0.06 ± 0.13 |
| B: without known events | 155 | +0.08 ± 0.10 | 52% | −0.01 | +0.05 ± 0.15 | +0.10 ± 0.13 |
| C: wide range only | 60 | +0.07 ± 0.15 | 52% | 0.00 | +0.12 ± 0.24 | +0.03 ± 0.18 |
| D: B and C (the system) | 55 | +0.09 ± 0.15 | 51% | 0.00 | +0.10 ± 0.26 | +0.09 ± 0.17 |
Three readings.
First: each filter improves the number a little, and none makes it significant. The step from A to D takes mean R from +0.05 to +0.09, but the error rises from ± 0.10 to ± 0.15 because trades fall from 169 to 55. All the estimates contain zero and also contain the random-direction result.
Second: the filters do not select better than a random choice. The 14 trades the calendar filter (B) removed had a mean R of −0.19 ± 0.45: it is pleasing, but the uncertainty is as large as the value. The 109 trades the energy filter (C) removed had a mean R of +0.05 ± 0.13: about the same as those kept. The Asian thermometer predicts the size of the day (chapter 3), not its direction.
Third: against random series the situation is this.
| Variant | Real | Random mean | 90% of random series | Share with R at least equal to the real one |
|---|---|---|---|---|
| A | +0.05 | −0.01 | from −0.11 to +0.10 | 18% |
| B | +0.08 | −0.01 | from −0.11 to +0.10 | 11% |
| C | +0.07 | +0.03 | from −0.09 to +0.16 | 30% |
| D | +0.09 | +0.03 | from −0.09 to +0.17 | 24% |
The most convincing is B, with 11% of random series reaching it: still not 5%. And among four variants the best must be judged on the maximum: in 300 random series the best of the four has an R equal to or above our +0.09 in 30% of cases. If you roll a die four times, 0.09 is not a surprising number.
Read the chart in words
H1 gold chart of 24 June 2026 with the Asian range, the short entry, the stop and the target.
| When | What | Price |
|---|---|---|
| 24 Jun 2026 00:00 UTC → 24 Jun 2026 06:00 UTC | Asian range | |
| 24 Jun 2026 11:00 UTC | short entry | 4,049.33 |
| 24 Jun 2026 11:00 UTC | stop | 4,114.97 |
| 24 Jun 2026 11:00 UTC | target 1R | 3,983.69 |
Times in UTC.
Read the chart in words
H1 gold chart of 7 August 2026 with the Asian range and the point where the base rule would have entered, marked as skipped by the system.
| When | What | Price |
|---|---|---|
| 7 Aug 2026 00:00 UTC → 7 Aug 2026 06:00 UTC | Asian range | |
| 7 Aug 2026 08:00 UTC | the system skips: first Friday | 4,311.63 |
Times in UTC.
The curve and how to read it
The cumulative R of the system (D) ends at +5.1 R over 55 trades, with a maximum drawdown of 1.7 R and a longest run of 4 consecutive losses. By month: December +0.2, January +2.7, February −0.3, March +0.6, April +0.1, May −0.5, June +2.5, July −0.8, August +0.1, September +0.4. Seven months out of ten are positive, but January and June alone are worth +5.2 R, more than the total. Without the two best months the system is at −0.1 R.
The three best trades are worth +1.00 R each (the cap: the target is at 1 R), so there is no single stroke of luck to remove. Nor is there a curious distribution of losses: the three worst are −0.76, −0.97 and −1.00 R. It is a system that makes small gains and small losses and no decisive trade.
Read the chart in words
Daily gold chart from December 2025 to October 2026 with, below, the cumulative R of the system and of random direction.
What four variants chosen beforehand say
One last check, the strictest. We pick the best variant on the first half of the sample and look at the second: the best in the first half is C (+0.12 R), which in the second half is worth +0.03 ± 0.18 R over 33 trades. That is, what looked like the advantage of selection disappears, or stays inside the error.
How long it would take to know
The standard deviation of the system's R is 0.56. To measure a mean R with an uncertainty of ± 0.10 about 120 trades are needed; for ± 0.05, about 477. System D makes on average 5.7 trades a month: 22 months are needed for ± 0.10 and seven years for ± 0.05. Every filter you add reduces the trades and lengthens the verification time: fewer trades, more years.
The honest question, then, is not "does it work?" but "how long am I willing to wait to know?". That is why in the previous books we insisted on the forward test: a rule fixed today and measured on the data that will come, without touching it, with a log. A six-month forward test of a system that makes 5.7 trades a month would give about 34 trades: still very few. To decide you need years, or a system that trades more often.
How to size it
Even if the result is not proven, the risk management part is worth it all the same. With a stop at the opposite extreme, the risk of a trade is the distance between entry and stop: in system D it has a median of 84 $ (from 67 $ at the 25th percentile to 142 $ at the 90th), and varies a lot from day to day. To risk 1% of a 10,000 $ account, that is 100 $, the size is the dollar risk divided by the stop distance: 100 $ divided by 84 $ makes about 1.2 ounces. If the stop is 142 $ (a very wide Asian range), the size falls to about 0.7 ounces. That is why risk is measured in R: the size adapts to the stop distance, and losing 1 R is always worth 1% of the account, whatever the number of dollars.
Read the chart in words
H1 gold chart of 20 January 2026 with the Asian range, the high and low lines and four candles marked with the letters A, B, C and D.
| When | What | Price |
|---|---|---|
| 20 Jan 2026 00:00 UTC → 20 Jan 2026 06:00 UTC | Asian range | |
| 20 Jan 2026 03:00 UTC | Point A | 4,680.49 |
| 20 Jan 2026 08:00 UTC | Point B | 4,731.48 |
| 20 Jan 2026 15:00 UTC | Point C | 4,750.97 |
| 20 Jan 2026 20:00 UTC | Point D | 4,759.01 |
Times in UTC.
What to do with it
- Do not put money on an unproven rule. The system gives +0.09 ± 0.15 R: the true value could be −0.06 or +0.24 R. A system with this error is not a working tool.
- Use it as a laboratory. The rules are clear, the data are public: you can measure them every new week, in a spreadsheet, without risking anything (demo or pen and paper).
- Fix the judgement criteria beforehand. For example: "after 150 trades, if mean R ± error still contains zero, I discard the rule". Without criteria you end up inventing them according to the result.
- Keep size on risk, not on price. 1% of the account per trade, size = risk divided by stop distance.
- The value of a badly proven system is in the questions it made you ask. What is random? How many things did I try? How does it do in the second half? These apply to any rule, even those you have not found yet.
Procedure in five steps
- Write the rule, the filters and the judgement criterion in advance.
- Test it on all available data, counting the variants you try.
- Compare with random direction, the two halves and random series; correct for the variants tried.
- Compute how long it would take to know (standard deviation, trades per month).
- If it is not proven, measure it forward without risking money, with a log, until the judgement criterion gives an answer.
Exercises
Chapter exercises
1. On the chart the Asian range (00:00-07:00 UTC) is in grey with its two lines. The system has neither stop nor target hit and closes on time at 21:00 UTC. Which letter is the last H1 candle before the time exit?
Read the chart in words
| When | What | Price |
|---|---|---|
| 20 Jan 2026 00:00 UTC → 20 Jan 2026 06:00 UTC | Asian range | |
| 20 Jan 2026 03:00 UTC | Point A | 4,680.49 |
| 20 Jan 2026 08:00 UTC | Point B | 4,731.48 |
| 20 Jan 2026 15:00 UTC | Point C | 4,750.97 |
| 20 Jan 2026 20:00 UTC | Point D | 4,759.01 |
Times in UTC.
2. The system gives +0.09 ± 0.15 R over 55 trades, random direction 0.00. What does it mean?
3. Among four variants the best has R = +0.09. In 300 random series the best of four reaches at least 0.09 in 30% of cases. What do you take from it?
4. The system has a standard deviation of R of 0.56 and makes 5.7 trades a month. How long does it take to measure R with ± 0.05?
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.