How many trades to test a strategy before your numbers mean anything

Start in plain English

How many trades does it take to test a strategy? Usually more than people expect, and more than a scalper's fast trade count makes it feel. This lesson explains why the first 20, 30 or even 100 trades of a plan leave a wide range of possible true results, why long losing streaks are normal even for a process with an edge, and why five trades on one stock in one morning are worth less than five trades on five different days.

The core idea: your results are an estimate, and an estimate from a small sample is wide. Three things decide how wide:

1. Sample size. The range around an average shrinks with the square root of the number of trades: four times the trades for half the width.
2. Spread of outcomes. Scalping outcomes range from about −1.3R to +2R per trade, so each trade carries a lot of noise relative to an edge of a few tenths of an R.
3. Independence. Trades that share a stock, a news event or a morning move together, so they add less information than their count suggests.

Two 30-trade runs from the same edge

All numbers here are Illustrative. Setup X is the fictional scalp from Thinking in probabilities: 40% of trades reach a +2R target, 45% hit a −1R stop and 15% scratch at 0R, with costs of 0.1R per trade. Its expectancy is +0.25R per trade after costs.

Two traders each run Setup X for 30 trades, following the plan exactly.

Review note: both traders ran the same +0.25R process. After 30 trades one has an average of +0.53R and the other −0.17R. Neither number is wrong; both are noisy estimates of the same thing. The rest of this lesson shows how to put a range around such an estimate, how often streaks like Path B's appear, and why counting correlated trades as independent makes the range look narrower than it is.

Variance and intervals in plain terms

An average from a sample comes with a range of plausible true values. A 95% interval is a range built so that, if you repeated the whole sampling process many times, about 95% of such ranges would contain the true value. It is a statement about the method, not a promise about your one range.

A rough interval for expectancy is: average ± 1.96 × (spread per trade ÷ √n), where the spread is the standard deviation of your per-trade results in R.

For the share of trades that reach the target, the Wilson interval works better than the simple formula when samples are small. If 12 of 30 trades reach the target (40%), the 95% Wilson interval runs from about 25% to 58%. At 40 of 100 it narrows to about 31% to 50%; at 120 of 300, about 35% to 46% (computed on these hypothetical counts).

Losing streaks are normal

Setup X loses 45% of its trades. Streaks of losses follow from that rate alone (computed by exact recursion, treating trades as independent; Illustrative):

A scalper who takes 300 trades in a few weeks should expect to see five losses in a row somewhere in them, even with a positive edge. A streak is not, on its own, evidence that the edge is gone. It becomes evidence when it comes with something else you can check: costs rising, fills worsening, the plan's conditions changing, or trades that no longer meet the written definition.

Correlated trades count for less

The formulas above assume each trade is an independent draw. Scalps often are not:

• Five scalps on the same stock in the same morning share one catalyst, one float and one crowd.
• Trades across stocks on the same day share the market's mood.
• A stock that runs for several days shows up in several sessions.

A rough way to see the effect is the effective sample size: n ÷ (1 + (m − 1) × ρ), where m is the number of trades per cluster and ρ is how strongly trades within a cluster move together (0 = independent, 1 = identical). Thirty trades taken five per morning on six mornings, with ρ = 0.3, carry about the information of 14 independent trades (Illustrative; ρ itself is hard to estimate).

FennEdge's own research faces the same issue. In the Gapper HOD study, 461 tickers produced 671 observations, because stocks that gapped on several days count once per session (2026-07-01 to 2026-10-02; FennEdge archive; as of 2026-10-02; see Small-cap gapper statistics (2026)). The study states this as a caveat: What time do small-cap gappers hit high of day?

When to stop testing

Decide the sample size before you start, and evaluate once at the end. Checking the result after every few trades and stopping as soon as it looks good, called optional stopping, makes a process with no edge look like it has one far more often than the interval suggests.

In a simulation of a zero-edge version of Setup X (20,000 runs of up to 100 trades, seed 20261005; computed, Illustrative), a single check at 100 trades showed a 95% range entirely above zero in about 2% of runs. Checking after every 10 trades and stopping at the first "pass" did so in about 11% of runs: more than four times as often, with no edge at all.

A practical protocol:

1. Write the definition, plan and sample size (for example 50 or 100 trades) before the first trade.
2. Log every trade that met the definition, including skipped ones, with costs.
3. Do not change the definition mid-block. If it needs to change, start a new block.
4. Evaluate at the end: expectancy, its interval, and whether the trades really were independent.

Observe vs. infer

After 30 trades, the trader in Path B reviews the log (Illustrative):

What shifts the odds

Here the "odds" are about how far your estimate can be trusted. These are statistical effects, not market findings:

• More trades narrow the range, slowly: four times the trades for half the width.
• Smaller spread per trade narrows it: fewer slipped stops and halt gaps mean less noise.
• Independent trades narrow it: spread across stocks and days.
• A fixed definition keeps the sample meaningful; changing it resets the count.
• A larger true edge is easier to detect: a +0.5R edge needs roughly a quarter of the trades a +0.25R edge needs.

Cost of being wrong

• Stopping too early on a bad run. The Path B trader who drops Setup X after 30 trades gives up +0.25R per trade of edge. At $50 per R and 30 trades a week, that is about $375 per week of expected value (Illustrative).
• Scaling up too early on a good run. If the true expectancy were zero and a trader doubled size after a Path A run, the doubled size would carry costs and risk with no edge behind it.
• Counting clustered trades as independent. A trader who sees "100 trades" when the information is closer to 40 trusts a range that is too narrow: the honest range is about 1.6 times as wide (√(100 ÷ 40) ≈ 1.6; Illustrative).

How many trades before this means anything

There is no single number. For a scalping edge of a few tenths of an R with a spread of about 1.4R per trade, a rough 95% range excludes zero only somewhere above 100 independent trades, and the estimate keeps tightening for hundreds more (Illustrative). A practical stance: treat the first 30 trades as checking that the plan can be executed and logged; treat 100 as a first estimate; and keep counting. The capstone, a measured 50-trade sim block, applies this to your own log.

What this lesson does not claim

• It does not claim any fixed number of trades proves an edge. The figures are computed from an Illustrative distribution.
• Normal-approximation intervals are rough for small samples and skewed outcomes such as halt gaps.
• Markets change. A sample from one period describes that period; the Advanced course covers regime changes.

Go deeper

• FennEdge first.
  • If you trade live, the Trading Journal syncs real fills from supported brokers, and its Review page (Pro and Max, as of 2026-10-05) lets you tag trades by setup, so you can count results per tag. It holds broker-synced trades only; practice and sim trades belong in the course practice log.
  • The Gapper HOD study is a worked example of stating sample size, repeat tickers and the period covered.
• Regime changes. A process measured in a busy small-cap month can behave differently in a quiet one. Splitting your log by month is a first check.
• Optional stopping in research. Armitage, McPherson and Rowe (1969) showed how repeated looks at accumulating data inflate false positives; the same applies to a trading log.
• The law of small numbers. Tversky and Kahneman (1971) describe how people expect small samples to look like the population they came from.
• Related: Expectancy, Risk/reward and breakeven win rate, A trading plan as a hypothesis, Reviewing scalps.

Exercises

Six exercises follow this lesson: three concept checks (what a 95% range means, what a losing streak does and does not show, why the sample size is fixed in advance), two calculations (a five-loss streak probability and the width of an expectancy range) and one scenario about which of two 40-trade samples carries more information. Exercises, answers and progress are available after signing in.

Primary sources

• Wilson, E. B. (1927). Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association 22(158).
• Armitage, P., McPherson, C. K. and Rowe, B. C. (1969). Repeated significance tests on accumulating data. Journal of the Royal Statistical Society, Series A 132(2).
• Tversky, A. and Kahneman, D. (1971). Belief in the law of small numbers. Psychological Bulletin 76(2).

Sources checked 2026-10-05. Setup X and all runs, intervals and streak figures are fictional or computed from an Illustrative distribution. Measured figures come from the FennEdge archive as stated. This lesson is education, not a recommendation to buy or sell any security. Past patterns do not indicate future results.