TL;DR
- There is no universal number of trades that makes a backtest reliable. The right count depends on the question and setup frequency. Sample variety matters too.
- For manual bar replay, use the first 60 observations as a structured discovery sample, not as proof that a strategy has an edge.
- Split those observations across trend and range conditions. Include high-volatility periods so one friendly market phase does not dominate the result.
- Decide your stop rules before the test. Pause and rewrite the setup when too many examples are unclear or the definition keeps changing.
- Move to fresh replay and paper trading only after the setup can be classified consistently without seeing the next candle.
Search for how many trades to backtest and you will find confident answers such as 30 or 50. Other guides argue for 100 or 200. Those numbers are easy to remember. They are much harder to defend without knowing what is being tested.
A coded strategy that produces one signal a month has a different sampling problem from a discretionary bar-replay drill that asks whether you recognized a pullback correctly. A sample drawn from one quiet quarter also says less than the same number of observations drawn from several market conditions. Counting trades is useful, but the count needs a job.
This guide gives that count a job. It uses a worked 60-observation plan for manual replay and a simple uncertainty table. Stop rules tell you when to revise the setup instead of collecting more of the same weak evidence.
How many trades should you backtest?
Start with the question you want the sample to answer. If the question is "Can I describe this setup clearly enough to find it again?", 20 carefully selected observations may expose obvious ambiguity. If the question is "Does this strategy have a stable win rate?", 20 is nowhere near enough. The second question needs a larger sample with realistic costs. It also needs out-of-sample checks and methods that account for dependence between trades.
Manual bar replay is often used for both jobs, which creates confusion. A trader may collect 60 replay trades and see a positive result. Calling the strategy validated would go beyond the evidence. Replay can help you test whether rules are usable and whether you can make the decision without future candles. It does not remove selection bias. Execution differences remain, and the chosen market period may have been unusually favorable.
The CFTC requires a prominent warning with presentations of hypothetical or simulated performance because those results have inherent limitations. The rule notes that simulated records do not represent actual trading and may understate or overstate the effect of market factors. That is a sound frame for a personal backtest too. A historical result is evidence about a defined sample, not a promise about the next trade.
A practical answer is therefore conditional:
- Use 20 observations to audit whether the setup definition is understandable.
- Use a 60-observation discovery sample to look for obvious condition dependence and recurring decision errors.
- Continue toward 100 or more only when the definition stays fixed and the extra observations add genuinely new conditions.
- Treat profitability or edge claims as a separate statistical project. Do not promote a replay checkpoint into proof.
What the sample size can and cannot tell you
Suppose you record each qualifying observation as a success or failure under fixed rules. A rough way to show sampling uncertainty is the worst-case 95% margin for a proportion: 1.96 multiplied by the square root of 0.25 divided by the sample size. This assumes independent observations and uses the most conservative proportion of 50%. It is an illustration, not a complete trading model.
| Observations | Approximate worst-case margin | Useful interpretation |
|---|---|---|
| 20 | ±21.9 percentage points | Definition audit only |
| 30 | ±17.9 percentage points | Early pattern check |
| 60 | ±12.7 percentage points | Structured discovery sample |
| 100 | ±9.8 percentage points | Still a broad estimate |
| 200 | ±6.9 percentage points | Narrower, but not automatically unbiased |
The calculation follows the standard large-sample approximation for a binomial proportion described in the NIST statistics handbook. It explains why a jump from 20 to 60 observations helps, while also showing why 60 is not a magic threshold. If 33 of 60 observations are winners, the observed rate is 55%. A rough interval can still span about 42% to 68% before allowing for market dependence. Fees and slippage add more uncertainty. Rule changes add another layer.
Trading observations rarely meet the clean independence assumption. Five trades taken during the same trend may share the same market driver. Repeated tuning creates another problem. If you test enough variations and retain only the best-looking one, the winning result may reflect selection rather than a durable edge. Research on backtest overfitting by David Bailey and coauthors examines this multiple-testing problem in detail.
Do not use the table to certify a strategy. Use it to keep the claim proportional to the evidence. A small sample can reveal a broken definition. It cannot establish a precise win rate.
A worked 60-observation bar-replay plan
Here is the ChartDojo sample-planning worksheet. The example uses one pullback setup on a single liquid market and timeframe. The plan is written before any result is viewed. That detail matters because a sample selected after seeing the outcome is already contaminated.
| Market condition | Target observations | What to record |
|---|---|---|
| Directional trend | 20 | Setup match, entry timing, invalidation level, result |
| Range or failed trend | 20 | Setup match, reason to skip, invalidation level, result |
| High volatility | 20 | Setup match, risk adjustment, invalidation level, result |
Choose historical windows with a rule that does not depend on how the chart eventually looked. One method is to list eligible months first, assign each month to a condition using a stated measure, then choose dates at fixed intervals. Keep a dated record of the selection rule. TradingView's Bar Replay documentation explains how to choose a starting point and move through historical bars without exposing the future chart.
For every observation, stop before the entry decision. Write one of three labels: qualifies, does not qualify, or unclear. If it qualifies, mark the planned entry and the price that invalidates the idea. Advance the replay only after the decision is fixed. Record the result, but grade the rule application separately from the outcome.
At observation 20, review clarity rather than profit. Count the unclear labels. Read the notes without looking at the result column. If the same phrase keeps appearing, such as "pullback was too deep" or "trend was not clean," the setup needs a measurable boundary.
At observation 40, compare the first two conditions. A setup that is easy to identify during trends may produce mostly unclear decisions in a range. Do not erase that difference by reporting one blended win rate. The condition split is information.
At observation 60, freeze the worksheet. Write a one-page summary with the classification rate and rule-adherence rate. Add the result by condition and the most common decision error. Then decide what the next test is. It may be another historical sample or a narrower setup definition. Fresh replay focused on execution is another option. More repetitions are useful only when they answer the remaining question.
Stop rules that protect the test
A weak backtest often continues because stopping feels like failure. In practice, an early stop can save hours and improve the next version of the test. Set these rules before the first replay session:
- Definition stop: pause if more than 4 of the first 20 observations are labeled unclear.
- Rule-change stop: restart the sample if the entry, invalidation, or setup filter changes after results are visible.
- Condition stop: add a new sampling block if one condition supplies more than half of all qualifying observations.
- Data stop: exclude a period only for a reason written before its result is known, such as missing bars or an unavailable session.
- Decision stop: move from testing the setup to practicing the trader when classification is consistent but timing or risk errors keep repeating.
The definition stop is not a universal statistical cutoff. It is a practice rule for this worksheet. Its purpose is to prevent the trader from quietly deciding what the setup means one chart at a time. If five of the first 20 examples are unclear, rewrite the definition and start a fresh sample. Do not relabel the old charts under the new rule and pretend they were always part of the plan.
The rule-change stop is stricter. A change can be sensible, but it creates a new test. Save the first result as version one. Write version two before opening a fresh historical window. This gives every result a visible lineage and makes hindsight harder to hide.
When to move from backtesting to paper trading
Move forward when the setup definition survives fresh examples, not when the historical profit looks exciting. You should be able to classify most observations without rewriting the rule. You should also know which market conditions remain weak and what decision must be made before the next candle appears.
Bar replay is the bridge between a written rule and a time-limited chart decision. The guide to using Bar Replay for trade practice explains how to hide future candles and structure the session. The comparison of free and paid trading practice tools can help you choose the next practice layer without paying for features that do not solve the current bottleneck.
Paper trading adds the market's clock and simulated order handling. It can test whether you wait for the setup during a normal session. It still does not recreate live fills or the emotional effect of real loss. Keep the claim narrow. Backtesting studies the rule. Replay rehearses the decision, while paper trading rehearses execution.
FAQ
Is 30 trades enough for a backtest?
Thirty observations may expose a vague setup or a basic data problem. It is usually too small for a precise estimate of performance. The approximate worst-case margin for a proportion at 30 observations is about 17.9 percentage points before accounting for dependence or selection bias.
Is 100 trades enough to prove a strategy works?
No fixed count proves that a strategy works. One hundred observations can narrow sampling uncertainty, but the result may still depend on one market regime or optimistic fills. Repeated tuning and omitted costs can weaken it further. Treat 100 as evidence inside a broader validation process.
Should I count skipped setups in the sample?
Record them, but do not mix them into the trade-result rate. A correct skip is valuable evidence about rule recognition. Keep separate counts for qualifying trades, correct skips, and unclear cases.
Can I mix several setups to reach the target faster?
Not if the goal is to assess one setup. Combining unrelated rules produces a larger number with a less useful meaning. Give each setup its own definition and sample. Keep the condition split separate too.
What if a setup appears too rarely to collect 60 observations?
Extend the historical window or reduce the claim. Do not loosen the setup after seeing that it is rare. Scarcity may be a real property of the idea, and it affects how quickly any estimate can improve.
Where ChartDojo fits
ChartDojo sits between paper trading and live trading. It is chart practice with AI-powered feedback on decision quality. The worksheet in this guide helps define and sample a setup. ChartDojo helps review whether the entry and risk decision made sense. It also checks the setup match before the outcome is known.
Use a simple daily cadence. Follow one setup and submit one decision. Review the feedback, then carry one correction into the next chart. Build real skills, not bad habits.
Start a free ChartDojo trial and use the first three sessions to review the most common error from your 60-observation worksheet.