Part 4: Checking and deciding · Chapter 12 of 14 · 10 min
How many results are needed
To see an edge of 0.1 R per trade, with a deviation of 1 R, you need about 385 trades for the error to equal the mean and about 785 to have an 80% chance of noticing it. At the pace of LA that is 14 months, at that of LD 20, at 5 trades a month 13 years. With 100 trades a true edge of 0.1 R is recognised only 17.5% of the time; and if you check every 25 trades a system with no edge looks "significant" at least once in 16% of paths. In our real series the lower limit of the interval is never above zero.
The question
You have a system and a series of results in R. When can you say it has an edge? Chapter 2 gave the answer in the form of an error: the observed mean has an uncertainty of ±1.96 × σ / √n. Here we take the next step: how many trades are needed for that uncertainty to be smaller than the edge you are looking for.
Two definitions:
- Distinguishable from zero: the lower limit of the interval (mean minus 1.96 errors) is above zero.
- Power: the probability that, if the edge is really there, you notice it. A power of 80% means that in one case out of five the true edge slips past you.
The formula
With a deviation σ (in R) and an edge μ you want to see, the number of trades is:
> for an error equal to the mean: n = (1.96 × σ / μ)² > for a power of 80%: n = 7.85 × σ² / μ²
For σ = 1 R (our deviation almost everywhere) and for σ = 0.56 R (that of SD):
| Edge μ | Error = mean (σ = 1) | Power 80% (σ = 1) | Error = mean (σ = 0.56) | Power 80% (σ = 0.56) |
|---|---|---|---|---|
| 0.05 R | 1,537 | 3,140 | 482 | 985 |
| 0.1 R | 385 | 785 | 121 | 247 |
| 0.2 R | 97 | 197 | 31 | 62 |
| 0.3 R | 43 | 88 | 14 | 28 |
| 0.5 R | 16 | 32 | 5 | 10 |
The number falls with the square of the edge: to see half the edge you need four times the trades. An edge of 0.05 R, small but common in real systems, requires more than 3,000 trades.
How long it takes
It depends on your pace. For our series (power 80%, μ = 0.1 R):
| Series | Trades per month | Trades needed | Months |
|---|---|---|---|
| LA | 57 | 777 | 14 |
| LD | 39 | 778 | 20 |
| SD (σ = 0.56) | 6 | 244 | 40 |
And for μ = 0.05 R: 4.5 years for LA, 6.7 for LD, 13.3 for SD. For a trader with 785 trades to make (μ = 0.1 R, σ = 1, power 80%):
| Trades per month | 5 | 10 | 20 | 60 | 100 |
|---|---|---|---|---|---|
| Years needed | 13.1 | 6.5 | 3.3 | 1.1 | 0.7 |
Whoever trades five times a month will not have, in a reasonable career, the statistics to tell an edge of 0.1 R from chance. They can decide with other tools (small risk, a plan with rules, a forward test), not with the statistics of their own account.
How often a true edge slips by
I simulate ±1 R systems with a given share of winning trades, and count in how many paths (3,000 for each cell) the lower limit of the mean exceeds zero:
| True edge | Winning share | 25 tr. | 50 | 100 | 250 | 500 | 1,000 |
|---|---|---|---|---|---|---|---|
| none (μ = 0) | 50% | 2.3% | 3.8% | 2.5% | 2.8% | 2.6% | 2.8% |
| μ = 0.05 R | 52.5% | 4.2% | 7.0% | 7.4% | 12.9% | 21.1% | 35.2% |
| μ = 0.10 R | 55% | 7.0% | 13.0% | 17.5% | 36.6% | 63.7% | 89.2% |
| μ = 0.20 R | 60% | 15.6% | 34.0% | 53.4% | 89.5% | 99.7% | 100% |
Three readings.
- An edge of 0.1 R at 100 trades is seen in 17.5% of cases. In the other 82 it tells you there is nothing.
- At 250 trades an edge of 0.2 R is almost always seen (89.5%), but one of 0.1 R only one time in three.
- Without an edge the false positive stays at 2.5-3% (the value expected with a single check). Small edges slip by, non-existent ones do not appear, if you check only once.
SX has a true edge of 0.21 R, and is a good example of how it shows. The lower limit of the interval exceeds zero for the first time at trade 59 (15 January 2026), and stays there steadily from trade 142. Before trade 58 (about a month) even a true edge is not distinguishable from chance, and until trade 141 (about 5 March 2026) the signal comes and goes. In the real series (LA, LD, SD) the lower limit is never above zero, at any moment.
Read the chart in words
Daily gold chart with, below, the curve of the mean R of SX and two dashed curves forming an interval that narrows over time.
Read the chart in words
Daily gold chart with, below, the curve of the mean R of LD and two dashed curves forming an interval that narrows over time and always includes zero.
Checking often deceives
The table above holds if you look at the result once, at the end. If instead you check every 25 trades and stop the first time the lower limit exceeds zero, the picture changes. Simulating 4,000 paths of a system with no edge, checking every 25 trades:
| Up to | Paths in which a "significant" result appears at least once |
|---|---|
| 250 trades (10 checks) | 10.4% |
| 500 trades (20 checks) | 12.5% |
| 1,000 trades (40 checks) | 15.9% |
With a single check it would be 2.5%. Checking all the time, about one in six systems with no edge passes the test at least once. It is a common mistake: "I waited for the result to become significant". The rule holds if the number of trades is decided beforehand.
The winning share has a larger error
If your target is the share of winning trades (p = 50%), the error (±1.96 × √(p × (1 − p) / n)) is:
| Trades | 25 | 50 | 100 | 250 | 500 | 1,000 |
|---|---|---|---|---|---|---|
| Error on the share | ±19.6 points | ±13.9 | ±9.8 | ±6.2 | ±4.4 | ±3.1 |
With 100 trades "a 55% success rate" means between 45% and 65%. That is why the winning share, alone, is a weak measure.
What to do with it
- Decide beforehand the number of trades (or the period) with which you will judge the system, and do not stop earlier because the result "looks good".
- Compute with the formula how many you need for the edge you expect, and how many months that is at your pace. If it is more than a year or two, your account alone is not enough.
- Until proof, treat the edge as zero in sizing (chapters 6 and 7): small risk, a cap on drawdown.
- Split the data: one part to build the system, one never-seen part to test it. The second part counts as a new test, with its own n.
- Accept the "inconclusive" result: it is not a failure, it is the honest answer with few trades.
Procedure in four steps
- Estimate the deviation of your results in R and the minimum edge that interests you.
- Compute n = 7.85 × σ² / μ² and divide it by the trades per month: those are the months of testing.
- Fix the rule in advance (number of trades, interval), and do not look at the outcome beforehand.
- At the end compare the lower limit of the interval with zero; if it does not exceed it, the edge is not demonstrated and the risk stays small.
Exercises
Chapter exercises
1. Deviation 1 R, edge sought 0.1 R. How many trades are needed for the error (±1.96·σ/√n) to equal the mean?
2. A system has a true edge of 0.1 R. With 100 trades, in how many cases does the lower limit exceed zero?
3. A system with no edge at all, checked every 25 trades up to 1,000. How many times, roughly, does at least one "significant" result appear?
4. At the pace of LA (57 trades per month), how many months are needed to distinguish with 80% power an edge of 0.1 R?
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.