How Many Trades for a Valid Backtest?
No trade count certifies a strategy. Choose the question and the precision you need, then examine whether the observations are independent and relevant to the market you intend to trade.
Why 100 trades is not a validity threshold
Thirty, 100 and 500 trades are counts, not statistical guarantees. A small set can expose a coding error. A much larger set can still contain future information, duplicated positions or rules selected using the same results.
The central limit theorem does not supply a universal 30- or 100-observation cutoff. Convergence depends on the distribution and assumptions. Decide whether the question concerns win probability, average net return, a ratio or a rare loss; each needs its own analysis.
A trade list also does not measure every risk. Mark-to-market losses while a position is open, missing delisted instruments and execution costs can matter even when the number of completed trades is large.
A reproducible win-rate example
Assume one strategy fixed in advance, independent win/loss outcomes and the same win probability for each trade. With 55 wins out of 100, the two-sided 95% Wilson interval using z = 1.96 is about 45.2% to 64.4%. The observed 55% is not a precise estimate.
For rough planning, solve E = z × sqrt(p(1 − p)/n). At p = 0.55 and an absolute margin E = 0.05, n = 1.96² × 0.55 × 0.45 / 0.05² = 380.3184, rounded up to 381. If the planning probability is unknown, p = 0.5 gives the largest variance and 385 observations under this approximation.
This calculation plans precision for a proportion. Detecting an improvement with specified statistical power is a different calculation. Neither answers whether the average win exceeds losses and trading costs. The NIST references below explain proportion intervals and test-design sample sizes.
Trade count, time coverage and dependence
Ten positions opened on the same market shock are not ten unrelated experiments. Overlapping holding periods, repeated entries and correlated symbols can share risk. Keep the unit of observation consistent; ten partial exits from one position should not silently become ten independent bets.
Retain timestamps, instrument, rule version and holding period. Review clusters and periods separately. A two-year run with many trades can miss conditions present in a longer, lower-frequency history. Neither trade count nor calendar length dominates in every case.
A 70% win rate does not imply low outcome variability: frequent small wins can coexist with occasional large losses. Avoid fixed trade-count thresholds for profit factor, Sharpe or Sortino. Record the return series, sampling interval and annualization convention before comparing those ratios.
What bootstrap simulation adds
Pineify Backtest Deep Report samples observed trade net P&L with replacement, adds those amounts to the starting capital and repeats the process. Its calculation defaults to 1,000 paths. A sampled path can duplicate one trade and omit another, so its final total can differ from the original.
Those paths are conditional on the observed pool and independent sampling model. They do not preserve clusters automatically or create an unobserved crash. More simulation runs improve numerical resolution under that model; they do not make the original sample larger. A narrow band from an unrepresentative pool can still be misleading.
The current report labels ruin using final equity at or below zero. A path that crosses zero and later recovers is not counted by that terminal test. Do not read it as a margin-liquidation or first-passage probability. Report percentiles are scenario summaries, not a certificate that a strategy has an edge.
An evaluation plan you can audit
- Freeze the rule and record selection. Save every tried variant, not only the winner. Repeatedly searching the same dataset changes how convincing a favorable result is.
- Reconcile the experiment. Check dates, costs, fills, duplicates and whether partial exits have been paired correctly. Inspect a few trades against the chart.
- Separate development from evaluation. Reserve later dates before tuning. A single holdout split is not the same as repeated walk-forward windows. A poor later result can reflect chance, costs or regime change as well as overfitting.
- Measure uncertainty for the chosen quantity. State the statistical model. If observations are dependent, independent-trade resampling may be unsuitable; a justified block or time-series method needs separate analysis.
- Stress what the history omits. Examine costs, gaps, concentration and open-position risk. Forward observation collects new evidence; it still cannot guarantee future returns.
Read report outputs as measurements
Backtest Deep Report analyzes an existing supported trade CSV. Its totals, distribution views and simulations help inspect that input. It does not rerun Pine code or establish that the source backtest avoided future information.
Increasing the count by optimizing for more entries can change the strategy and add selection bias. Pooling instruments can also create correlated exposures. Keep the research question fixed and report unresolved uncertainty when relevant new observations are unavailable.
For a fixed equity-path prefix, extending that same path cannot reduce its already-observed maximum peak-to-trough drawdown. A changed date window, starting capital or sizing rule is a different comparison. None of these observed maxima bounds future losses.
Sources
- NIST: Wilson and exact proportion confidence intervals
- NIST: sample sizes for tests of proportions
- SciPy: bootstrap sampling and confidence-interval assumptions
Examples are hypothetical statistical calculations, not measured strategy returns. Source review: October 20, 2026.
FAQ
Inspect Your Backtest Trade Sample
Review a supported trade CSV, compare its distributions and examine bootstrap scenarios. Statistical validity still depends on the data, research design and assumptions you verify.
Analyze My Backtest Now