Answer and evidence
How uncertain is an annualised Sharpe ratio estimated from 24 monthly observations?
Short answer. With 24 monthly observations, an observed annualised Sharpe ratio of 1.00 carries an asymptotic 95% confidence interval of −0.41 to +2.41 (±1.41) under independent normal returns. Supporting analysis →
- Evidence type
- Evidence review; a review or worked illustration, not a new causal experiment.
- Sample and scope
- 2026-11-04. Asymptotic 95% confidence intervals for annualised Sharpe ratios estimated from N monthly observations (Lo 2002, IID normal baseline). Published measurements and sample counts.
- Limits of this answer
- Asymptotic normal approximation assuming independent and identically distributed monthly returns. Fat tails, skewness, autocorrelation and data-mining across multiple parameter sets widen true uncertainty.
Key findings
- To exclude a zero true Sharpe ratio at the two-sided 95% confidence level over a 24-month sample requires an observed annualised Sharpe ratio of at least 1.45. Evidence →
- Non-normal fat tails, positive skewness bursts, serial correlation and multiple backtest trials widen confidence intervals beyond standard textbook formulas. Evidence →
Research reviewed 2026-11-04; page presentation updated 2026-11-04. These findings describe the studied market or measurement method, not any individual strategy’s rules.
Published evidence
| Observed annualised Sharpe | 12 months (95% CI) | 24 months (95% CI) | 36 months (95% CI) | 60 months (95% CI) |
|---|---|---|---|---|
| 0.50 | −1.47 to +2.47 (±1.97) | −0.89 to +1.89 (±1.39) | −0.64 to +1.64 (±1.14) | −0.38 to +1.38 (±0.88) |
| 1.00 | −1.00 to +3.00 (±2.00) | −0.41 to +2.41 (±1.41) | −0.15 to +2.15 (±1.15) | +0.11 to +1.89 (±0.89) |
| 1.50 | −0.55 to +3.55 (±2.05) | +0.05 to +2.95 (±1.45) | +0.32 to +2.68 (±1.18) | +0.58 to +2.42 (±0.92) |
| 2.00 | −0.12 to +4.12 (±2.12) | +0.50 to +3.50 (±1.50) | +0.78 to +3.22 (±1.22) | +1.05 to +2.95 (±0.95) |
| 3.00 | +0.70 to +5.30 (±2.30) | +1.37 to +4.63 (±1.63) | +1.67 to +4.33 (±1.33) | +1.97 to +4.03 (±1.03) |
Why a two-year Sharpe ratio has wide error bars
The Sharpe ratio divides mean return by return volatility, both of which must be estimated from finite data. For N independent, normally distributed monthly observations with non-annualised Sharpe SR_m = SR_ann / √12, Andrew Lo (2002) showed that the asymptotic standard error of the annualised Sharpe ratio is √((1 + 0.5 × SR_m²) / N) × √12.
With N = 24 months, the standard error is roughly 0.71 to 0.83 annualised Sharpe units. Multiplying by 1.96 gives a 95% confidence half-width of ±1.39 to ±1.63. An observed Sharpe of 1.00 over 24 months is therefore statistically compatible with a true Sharpe anywhere between −0.41 and +2.41.
How many months are needed to distinguish skill from noise?
Under the baseline formula, the minimum observed annualised Sharpe ratio needed for the lower bound of the 95% confidence interval to exceed zero is the value of SR that solves SR = 1.96 × SE(SR), slightly above the simple approximation 1.96 × √(12 / N). That threshold is 2.14 at 12 months, 1.45 at 24 months, 1.16 at 36 months and 0.89 at 60 months.
Counting daily or hourly bars in the same two-year window does not shrink the standard error of the mean return. Higher-frequency sampling improves the volatility estimate, but the uncertainty of the average return depends on the total time span T in years (SE ≈ 1 / √T), not the number of subdivisions within T.
What makes crypto Sharpe ratios even noisier
Strategies that remain flat in quiet months and trade only around specific market events have fewer active observations than total calendar months. If only 8 of 24 months contain trades, the effective sample size for estimating the return distribution is closer to 8 than 24.
When only one or two months in a sample are negative, downside-only ratios such as Sortino or Calmar become unstable and can jump into the tens or hundreds from a single small loss. Selection bias across multiple backtest variations further inflates the best-looking Sharpe ratio relative to out-of-sample expectation.
How is the standard error of an annualised Sharpe ratio calculated?
For N monthly observations and observed annualised Sharpe SR, convert to monthly SR_m = SR / √12, compute the monthly standard error √((1 + 0.5 × SR_m²) / N), and multiply by √12. The 95% confidence interval is SR ± 1.96 × SE.
Does using daily data instead of monthly data narrow the Sharpe confidence interval?
Almost not at all for the mean return component. Because the mean and standard deviation scale differently with sampling frequency, the standard error of the annualised Sharpe ratio depends primarily on the number of years of history rather than whether returns are measured daily or monthly.
Limitations
Asymptotic normal approximation assuming independent and identically distributed monthly returns. Fat tails, skewness, autocorrelation and data-mining across multiple parameter sets widen true uncertainty.