We may earn a commission through these partner links. Demo funds are virtual; access and country eligibility depend on the platform terms. Disclosure
A trading win rate needs a numerator, denominator and collection method. “80% accuracy” could mean four wins in five observations or 800 in 1,000. The percentages match, but the uncertainty differs under the same statistical assumptions. Neither sample size alone nor an impressive winning streak proves a reliable edge.
Research checked 23 October 2026. Statistical examples are illustrative and assume conditions that must be assessed separately for trading data.
Always report wins and observations together
An observed win rate is wins divided by a declared denominator. State whether it counts accepted contracts, decided outcomes, all entries or generated signals. These are different populations. A signal that was never entered is not an accepted transaction.
Ten wins, ten losses and five full refunds give 50% wins among decided outcomes or 40% among all entries. Both can be correct if labelled. Preserve refunds separately and include their actual financial treatment in the ledger.
A claim of “20 wins” also needs the loss count. Twenty wins from 25 observations and twenty from 50 have very different rates. A number of wins without total observations is incomplete evidence.
20 wins ÷ 25 observations = 80%
20 wins ÷ 50 observations = 40%
Invented sample counts; neither percentage proves a future win probability.
How sample size affects uncertainty
Under an independent, stable Bernoulli model, each observation has the same win probability and a win/loss outcome. An illustrative estimated standard error of a sample rate p-hat is sqrt[p-hat × (1 − p-hat) ÷ n].
| Observed wins / n | Sample rate | Approximate estimated standard error |
|---|---|---|
| 12 / 20 | 60% | 10.95 percentage points |
| 60 / 100 | 60% | 4.90 percentage points |
| 300 / 500 | 60% | 2.19 percentage points |
| 600 / 1,000 | 60% | 1.55 percentage points |
This table illustrates sampling variability, not a confidence interval and not proof that the model fits the observations. Increasing n fourfold halves this estimated standard error when the observed rate stays the same.
At rates near zero or one, the simple plug-in expression can be misleading. Twenty wins out of twenty do not establish a known 100% probability simply because that expression becomes zero. An interval method appropriate to the model is needed.
Confidence intervals need a method and assumptions
NIST’s binomial-proportion interval guidance discusses methods including Wilson intervals. It is a useful reference for distinguishing a point estimate from an uncertainty interval. A software output should identify its method and confidence level.
A 95% frequentist procedure is interpreted through its coverage over repeated comparable samples under the model. It should not be casually described as a 95% probability that this fixed unknown rate lies in the one interval already calculated.
More importantly, an interval for a stable binomial model does not automatically account for changing sessions, correlated entries or selection of the best-performing rule. Those are separate evidence problems.
Why 100 trades may contain less information than expected
Several contracts opened simultaneously on the same symbol can depend on nearly identical price movement. Related symbols or overlapping expiries can also share drivers. Counting each entry does not make those observations independent.
Retain timestamps and group identifiers so the relationship can be reviewed. Do not divide an estimated error by the square root of a large raw count while ignoring strong overlap. The appropriate statistical treatment depends on the actual dependence structure.
Duplicated screenshots, copied signals and repeated records are another issue. Deduplicate observations using stable identifiers before reporting sample size. Keep the exclusion log so cleaning does not quietly remove genuine losing cases.
Changing rules means changing the measured process
If indicator settings, entry conditions or exclusions change, record a new rule version. A combined sample across versions estimates a mixture, not necessarily the performance of the final version now being advertised.
The same applies to feed, symbol, timeframe, expiry and account environment. A daily-stock study and an M1 OTC study should not be pooled merely because both use the same candle-pattern name.
Report the distribution across sessions too. A large count gathered in one brief market episode may say little about other conditions. Observation count and coverage of conditions are different dimensions of evidence.
A large selected sample can still mislead
TradingView’s strategy documentation discusses selection bias, overfitting and future-data leakage. Those problems are relevant when interpreting manual or automated studies, even though a particular broker contract needs its own data and payoff model.
If 50 rule variants are tried and only the best is published, the winning sample reflects that search. Disclose alternatives and unsuccessful versions. Repeatedly checking the evaluation period and tuning again turns that period into development data.
Likewise, stopping a public record immediately after a winning streak can select a favourable endpoint. Predefine review dates or observation rules and preserve the full sequence rather than reporting whichever slice looks best.
Win-rate uncertainty is not the whole profitability question
For a simple 80% net payout with full-stake losses and no other outcomes, the break-even probability is approximately 55.56%. An observed 60% sample rate is above that numerical threshold, but uncertainty and model fit still need evaluation.
Changing payout to 60% makes the threshold 62.5%. The same 60% rate then produces negative net results in the equal-stake model. Preserve accepted payouts and stakes instead of treating accuracy alone as profitability.
Variable stakes, fees, refunds and adjustments require transaction-level accounting. A statistical claim about direction does not establish execution quality or withdrawal reliability.
A sample-quality reporting checklist
- Count: wins, losses, refunds, exclusions and missing outcomes.
- Coverage: dates, sessions, full symbols and providers.
- Definition: frozen entry, outcome and exclusion rules.
- Version history: changes and alternatives tested.
- Timing: information available at decision and accepted execution.
- Economics: actual stakes, returns and net results.
- Dependence: overlap, duplicate records and shared events.
- Evaluation: a period not used to select the final rule.
A clear report can say “12 wins and eight losses under V1 in this defined session” without claiming a certified edge. Precision about what was measured is more useful than extra decimal places on an unsupported forecast.
Frequently asked questions
Are 100 or 500 trades enough?
There is no universal sufficient count. Required evidence depends on uncertainty, the claimed effect, dependence, selection and process stability.
Does a larger sample fix cherry-picking?
No. More selected observations can preserve the same selection problem.
Do twenty consecutive wins prove reliability?
No. The complete sampling and stopping rules, other outcomes and payoff record still matter.
Sources and further learning
NIST supports the attributed interval concepts and TradingView the testing-bias discussion. Calculations and checklist are educational analysis. See our calculation tools, demo practice guide and methodology.
Risk: Binary contracts can lose the full stake. A historical sample rate does not guarantee future performance.