How Long Should You Backtest a Trading Strategy?
A longer backtest helps, but the research search can give the noise more chances to win
A five- or ten-year minimum sounds disciplined until you ask how many strategy variants were tried. Under a transparent zero-edge benchmark, 100 independent trials with one year of daily data produce a median winning Sharpe of 2.48. Five years cuts that median to 1.10 and ten years to 0.78. But if the ten-year search expands to 1,000 independent trials, the median winner rises back to 1.01.
What pure search luck looks like as the backtest gets longer
Start with 100 independent strategies that all have a true expected return of zero. Select only the strategy with the highest annualized sample Sharpe. With 252 daily observations, the median winner is about 2.48. That result is not a simulation accident. Under the stated Normal null, the sampling distribution is known exactly, so the selected maximum can be calculated directly from a Student-t distribution.
Longer samples reduce false winners, but broader searches push them back up
Across the exact zero-edge benchmark, longer samples lower the median winning Sharpe at every search breadth. With 100 trials the median winner falls from 2.48 after one year to 0.78 after ten years, while expanding the ten-year search to 1,000 trials raises the median winner to 1.01.
| Backtest length | Trials | Median best Sharpe | P(best Sharpe ≥ 1) | P(best Sharpe ≥ 2) |
|---|---|---|---|---|
| 1 year | 1 | 0.00 | 15.9% | 2.3% |
| 1 year | 100 | 2.48 | 100.0% | 90.5% |
| 1 year | 1,000 | 3.23 | 100.0% | 100.0% |
| 2 years | 1 | 0.00 | 7.9% | 0.2% |
| 2 years | 100 | 1.75 | 100.0% | 21.6% |
| 2 years | 1,000 | 2.27 | 100.0% | 91.2% |
| 3 years | 1 | 0.00 | 4.2% | 0.028% |
| 3 years | 100 | 1.42 | 98.6% | 2.8% |
| 3 years | 1,000 | 1.85 | 100.0% | 24.5% |
| 5 years | 1 | 0.00 | 1.3% | 0.000% |
| 5 years | 100 | 1.10 | 72.3% | 0.042% |
| 5 years | 1,000 | 1.43 | 100.0% | 0.4% |
| 10 years | 1 | 0.00 | 0.1% | 0.000% |
| 10 years | 100 | 0.78 | 7.6% | 0.000% |
| 10 years | 1,000 | 1.01 | 54.7% | 0.000% |
| 20 years | 1 | 0.00 | 0.0% | 0.000% |
| 20 years | 100 | 0.55 | 0.0% | 0.000% |
| 20 years | 1,000 | 0.72 | 0.4% | 0.000% |
The contrast at ten years is the part I would keep in mind. Search 100 independent zero-edge variants and the median winner is about 0.78. Search 1,000 and it is about 1.01. The historical sample did not get shorter. The research process simply received more chances to select an extreme result.
Why “use at least 10 years” is not a statistical answer
Calendar length is useful for at least two reasons. More observations narrow sampling uncertainty, and a longer history can expose a strategy to more market environments. Neither benefit creates a universal cutoff where the backtest becomes trustworthy.
The missing variable is the research search itself. A researcher can test one pre-specified rule, or thousands of combinations of indicators, lookbacks, thresholds, universes, rebalance frequencies, filters, and exits against the same ten years of history. Those are not the same experiment. The Deflated Sharpe Ratio work by David H. Bailey and Marcos López de Prado formalized the selection problem: a reported winner should not be evaluated as though it were the only idea that was tried. A 2026 Sharpe-ratio inference paper separately emphasizes sample-length requirements and multiple testing among the recurring errors in performance inference.
This is also why “number of trades” and “number of years” cannot substitute for one another. A high-frequency strategy can accumulate many trades inside one market regime. A low-turnover strategy can span several regimes with relatively few decisions. Both sample size and economic coverage matter, and neither reveals how heavily the researcher searched the data.
The benchmark is an exact order-statistic calculation
Assume each strategy has independent, Normally distributed daily returns with true mean zero. For one strategy with n observations, the usual t-statistic for the sample mean follows a Student-t distribution with n − 1 degrees of freedom. Annualizing the sample Sharpe with 252 observations per year gives:
If F is the Student-t cumulative distribution and M independent zero-edge strategies are tested, the probability that every strategy stays below annualized Sharpe s is:
The median winning Sharpe is therefore the single-strategy Sharpe quantile associated with probability 0.51/M. The downloadable grid evaluates that expression for one, two, three, five, ten, and twenty years of daily observations and search sizes of 1, 10, 100, and 1,000 independent strategies.
The point of using an exact benchmark is transparency. A reader can reproduce every table cell without choosing a random seed or trusting a large Monte Carlo run.
The assumptions are intentionally stricter than reality in some places and simpler in others
Real strategy searches are not collections of mutually independent bets. Nearby parameter choices and related indicators are often strongly correlated. That means 1,000 literal backtests should not automatically be interpreted as 1,000 independent trials. Grizzly Bulls has a separate correlated-strategy-search benchmark for that issue.
Real returns are also not reliably IID Normal. Serial correlation can reduce the amount of independent information in a track record, while skewness, heavy tails, changing volatility, and regime shifts alter Sharpe-ratio inference. The autocorrelation calculator and Probabilistic Sharpe Ratio calculator address pieces of those problems.
Most importantly, this table cannot detect look-ahead bias, survivorship bias, unrealistic trading costs, data leakage, parameter instability, or repeated reuse of an out-of-sample period. Twenty years of contaminated data can be worse evidence than a shorter but properly designed test.
What to test after deciding the sample is long enough to be informative
I would treat sample length as one part of a validation stack rather than a publication threshold. The most useful next questions are whether realistic costs consume the edge, how much multiple testing occurred, whether the selected Sharpe remains unusual after that search, and whether the result survives genuinely untouched data.
Download the benchmark data
The public files contain all 24 combinations of backtest length and independent strategy count, including the median winning annualized Sharpe and the probability that the selected winner reaches Sharpe 1.0, 1.5, or 2.0.
Research data
- Download CSVCSVCSV downloadTabular public study data for spreadsheet analysis, independent checks, and new charts.Data snapshot September 17, 2026 · Reuse with attribution to the canonical study.
- Download JSONJSONJSON downloadStructured public study data for programmatic verification while preserving the published field names and research context.Data snapshot September 17, 2026 · Reuse with attribution to the canonical study.
Sources and related methodology
- Bailey & López de Prado (2014), “The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality”
- López de Prado, Lipton & Zoonekynd (2026), “Sharpe Ratio Inference: A New Standard for Decision Making and Reporting”
- López de Prado & Porcu (2026), “The Deflated Sharpe Ratio: A Unified Framework for Search-Adjusted Performance Inference”
When citing the benchmark, preserve both the backtest length and the number of independent strategy trials. The result is a zero-edge null benchmark, not a claim about the false-positive rate of a specific real strategy search.