Part 3: Time, stop and size · Chapter 10 of 14 · 9 min
Volatility and a rule's result
Cutting the plan's 79 trades into three by ATR at entry gives +0.29, +0.90 and +0.80 R (no order); cutting them by the ATR's position over the last 500 candles gives +0.20, +0.55 and +1.32 R (an apparent order). Across all 480 rules of the grid, keeping only high ATR changes the mean R by −0.024 (209 rules out of 480 improve), against −0.019 for random series: volatility does not improve the rules in this data, and the cut halves the trades.
The question
Does a rule work better when the market is agitated or when it is calm? It is one of the most frequent questions, and one of the most dangerous, because it has many possible answers: volatility can be measured in many ways, divided into many groups and, among all the combinations, one will almost always give a result that looks good. Book 12 calls this researcher degrees of freedom.
To answer honestly you need a rule: splits are written down before looking at the results and all of them are counted. Here we do it by showing three splits of the same 79 trades of the example plan and then the same question put to 480 rules.
Three cuts of the same 79 trades
Cut 1: ATR at entry, in dollars. The terciles of the hourly ATR are $16.5 and $21.3.
| ATR at entry | Trades | Mean R ± error | Sum | Winning |
|---|---|---|---|---|
| Low (up to $16.5) | 17 | +0.29 ± 0.89 | +5.0 R | 35% |
| Middle | 35 | +0.90 ± 0.64 | +31.4 R | 51% |
| High (above $21.3) | 27 | +0.80 ± 0.76 | +21.7 R | 44% |
Here the low-volatility group does worse (+0.29), but the other two are almost equal: no order.
Read the chart in words
Daily gold chart with, below, three cumulative R curves: the grey one rises little, the blue and pink ones much more.
Cut 2: position of the ATR over the last 500 candles. Instead of the dollar value we look at whether the ATR of that moment is low or high compared with the last 20 days or so (terciles of the position).
| ATR relative to the last 500 candles | Trades | Mean R ± error |
|---|---|---|
| In the low third | 23 | +0.20 ± 0.77 |
| In the middle third | 27 | +0.55 ± 0.71 |
| In the high third | 29 | +1.32 ± 0.71 |
Here the order is clear: the higher the ATR relative to the recent past, the better the rule does. The difference between the high and the low third is 1.12 R; the combined error is about ±1.05. It sits at the limit.
Cut 3 (chapter 9): only entries with ATR above the 80th percentile. 18 trades, +1.44 ± 0.93 R against +0.53 ± 0.47 for the other 61.
Why three cuts are not enough
Three splits of the same series gave three different pictures: the first without order, the second and third with an order that rises with volatility. Which is «true»? With 79 trades the errors are ±0.7 R in each group: none of the three pictures can be told from the others with confidence. And we looked at three splits and reported them all: in real research ten would be looked at and the best shown.
The correct method is to ask whether the same question put to other rules gives the same result. If keeping only high ATR really improved breakout rules, it should do so for many rules, not for one.
The same question for 480 rules
We took the grid of 480 rules from book 12 (breakout and fade, five lengths, four stops, four targets, three durations) and for each we calculated the mean R with all entries and with only entries in high ATR (above the upper tercile) and in low ATR (below the lower tercile).
| All entries | High ATR only | Low ATR only | |
|---|---|---|---|
| Grid mean R | −0.017 | −0.042 | −0.023 |
| Average trades per rule | 226 | 87 | 74 |
| Mean difference in mean R (against all) | – | −0.024 | −0.006 |
| Rules that improve | – | 209 out of 480 (44%) | 233 out of 480 (49%) |
Keeping only high ATR worsens things on average (−0.024) and improves fewer than half of the rules. To know what chance would give, we repeated the same experiment on 40 random series (candles shuffled in one-day blocks, without trend):
| Mean difference in mean R | |
|---|---|
| Random series, high ATR only | −0.019 (from −0.043 to +0.008 in 90% of cases) |
| Random series, low ATR only | +0.010 (from −0.031 to +0.055) |
The real values (−0.024 and −0.006) fall inside the band of chance. There is no effect of volatility on the breakout and fade rules of this grid. The filter has only one sure effect: it cuts the trades to a little more than a third (226 against 87), which widens the error by about 60% (proportional to 1/√n).
For the best rule of the grid (the 80-candle breakout, our plan) the picture is that of cut 1: +0.74 R with all 79 trades, +0.58 with 38 in high ATR, +0.54 with 22 in low ATR. Both filters remove entries and the mean R stays in the same order of magnitude.
What follows
- A volatility filter is not a good idea by default: on the grid it does not improve things; it reduces trades.
- A rule may also behave differently at low and high levels, but it needs a proof written beforehand and checked out of sample (book 12).
- An apparent order across three groups of 20-30 trades is a hypothesis, not a result.
- The number of splits tried is declared: here it is three on the plan and two on the grid.
What to do
- Write the volatility split first that you want to test (which measure, how many groups, which thresholds).
- Report all the splits tried, not just the one that «works».
- Ask whether it also works on other rules or on a second part of the data.
- Compare with random series (book 12): if chance gives the same effect, there is no effect.
- Do not cut trades without a written reason: every filter widens the error.
Procedure in four steps
- Write the volatility split and the thresholds before looking at the results.
- Calculate the mean R with its error in each group.
- Repeat the same split on other rules and on random series.
- Keep the filter only if the effect sits outside the band of chance and holds in a second part of the data.
Exercises
Chapter exercises
1. Splitting the 79 trades by ATR position, the high third makes +1.32 ± 0.71 R, the low +0.20 ± 0.77. Why is it not proof?
2. Across 480 rules, keeping only entries with high ATR changes the mean R by −0.024 and on random series by −0.019. What does it mean?
3. The high-ATR filter takes the average trades per rule from 226 to 87. What is the statistical effect?
4. When can a volatility filter be accepted?
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.