The Standard Error of the Sharpe Ratio

Someone shows you a strategy with a Sharpe ratio of 1.5 over three years. Is that good evidence of skill, or could a mediocre strategy have produced that number by luck? Almost every argument about backtests is really this argument, and it is settled by a single quantity: the standard error of the Sharpe ratio.

The instinct to resist is treating the reported number as a fact. A Sharpe ratio calculated from data is a statistic, computed from a finite sample, and if you re-ran the same strategy through a different three years you would get a different number. How different is the whole question.

The formula

For $T$ return observations that are independent and identically distributed, the estimated Sharpe ratio $\widehat{SR}$ has approximate standard error

$\text{SE}\big(\widehat{SR}\big) \;\approx\; \sqrt{\frac{1 + \widehat{SR}^{\,2}/2}{T}}.$

In words: the uncertainty shrinks with the square root of how many observations you have, and grows slightly when the Sharpe ratio itself is large. Both $\widehat{SR}$ and $T$ must be measured at the same frequency, if you use daily returns then $\widehat{SR}$ is the daily Sharpe ratio and $T$ is the number of days.

That frequency-matching is where people go wrong, so here is the version you should actually memorise. Convert to annual terms and almost everything cancels, leaving

$\text{SE}\big(\widehat{SR}_{\text{annual}}\big) \;\approx\; \frac{1}{\sqrt{N}}, \qquad N = \text{years of data}.$

In words: the error bar on an annualised Sharpe ratio is one over the square root of the number of years, and nothing else. Sampling more finely does not help. Ten years of daily data and ten years of monthly data give you almost exactly the same certainty about the Sharpe ratio, because what you are trying to measure is an annual quantity and you only have ten years of it.

Because the estimate is approximately normal around the truth, you can see the whole picture as a bell curve. Set the mean to a plausible true Sharpe ratio and the standard deviation to $1/\sqrt{N}$, then read off how much probability sits below zero, that is the chance a genuinely good strategy shows a losing three-year stretch:

Worked example: is a 1.5 Sharpe over three years real?

$N = 3$ years of daily returns, $\widehat{SR} = 1.5$ annualised.

Step 1, standard error. $\text{SE} = 1/\sqrt{3} = 0.577$. (The exact formula gives 0.579, the difference is negligible.)

Step 2, test against zero. $t = 1.5/0.577 = 2.60$. That clears the usual 1.96 threshold, so yes, the record is statistically significant at the 5% level.

Step 3, and now the uncomfortable part. The 95% confidence interval is

$1.5 \pm 1.96 \times 0.577 = [\,0.37,\ 2.63\,].$

The data is consistent with a strategy whose true Sharpe ratio is 0.4, barely worth running after costs, and equally consistent with one of 2.6, which would be among the best in the world. "Significant" and "well estimated" are not the same thing. You have ruled out zero and almost nothing else.

Step 4, one year instead of three. $\text{SE} = 1.0$, so $t = 1.5$ and the interval is $[-0.46,\ 3.46]$. A single year of a 1.5 Sharpe ratio is not evidence of anything.

Worked example: how long a record do you need?

Rearrange the significance condition $\widehat{SR}/\text{SE} > 1.96$ with $\text{SE} = 1/\sqrt{N}$:

$N > \left(\frac{1.96}{SR}\right)^{2} \ \text{years}.$

In words: required track record equals about four divided by the Sharpe ratio squared. Running the numbers:

True Sharpe	Years needed
0.3	43
0.5	15
1.0	3.8
1.5	1.7
2.0	1.0

This single table explains a great deal about the industry. A long-only equity manager with a genuine 0.3 edge cannot prove it within a career. A market-neutral book at 1.0 needs four clean years. Only high-Sharpe, high-frequency strategies can be validated quickly, which is exactly why those desks measure everything and long-horizon investors argue forever.

Fat tails make it worse

The formula above assumes returns are normal. Real strategies, especially ones that sell insurance, collect carry, or short volatility, have negative skew and heavy tails. Mertens' correction adds those moments in:

$\text{Var}\big(\widehat{SR}\big) = \frac{1 + \tfrac12 SR^2 - \gamma_3\,SR + \tfrac{\gamma_4 - 3}{4}\,SR^2}{T},$

where $\gamma_3$ is skewness and $\gamma_4$ is kurtosis. In words: negative skew (the $-\gamma_3 SR$ term, which becomes positive when skew is negative) and fat tails (the $\gamma_4 - 3$ term) both inflate the error bar. The strategies whose Sharpe ratios look most impressive are frequently the ones whose Sharpe ratios are least reliable.

A concrete case. Five years of monthly returns, so $T = 60$. Monthly Sharpe ratio 0.30, which annualises to $0.30\sqrt{12} = 1.04$. Skewness $-1.5$, kurtosis 8.

• Naive: $\text{Var} = (1 + 0.045)/60 = 0.01742$, so monthly SE $= 0.132$, annualised $0.457$. Then $t = 1.04/0.457 = 2.27$, significant.
• Corrected: $\text{Var} = (1 + 0.045 + 0.450 + 0.113)/60 = 0.02679$, so monthly SE $= 0.164$, annualised $0.567$. Then $t = 1.04/0.567 = 1.83$, not significant.

The skew and kurtosis terms contributed 0.563 of the 1.608 total, more than a third of the variance, and flipped the verdict.

In practice

• Report the interval, not the point. "Sharpe 1.5" invites false confidence; "Sharpe 1.5, 95% CI 0.4 to 2.6 on three years" is honest and takes no more space.
• Compare two strategies carefully. The difference of two Sharpe ratios has its own, wider standard error. A 1.4 and a 1.1 over the same five years are statistically indistinguishable.
• Use it as a design constraint. Before building anything, ask how many years of data you have and what Sharpe ratio would be detectable in that span. If the honest answer is 0.9 and your idea plausibly delivers 0.3, no amount of modelling will settle the question.
• Watch the assumptions. Autocorrelated returns, common in illiquid or smoothed marks, break the i.i.d. assumption in the other direction and understate risk. See Autocorrelation and Serial Correlation and Sharpe Ratio.