The Deflated Sharpe Ratio

The Deflated Sharpe Ratio (DSR) answers the question every impressive backtest raises: is this Sharpe ratio real, or is it the luckiest of many tries? An observed Sharpe of 2 means one thing if it is the only strategy you ever tested and something entirely different if it is the best of a thousand. Bailey & López de Prado's DSR is a single number, a probability, that corrects the observed Sharpe for three things ordinary Sharpe ignores: the number of trials, the non-normality (skew and kurtosis) of the returns, and the length of the track record. It is the rigorous form of the intuition behind Data-Snooping Bias.

Step 1: the Probabilistic Sharpe Ratio

Start with the sampling uncertainty of the Sharpe estimate itself. The estimated (per-observation) Sharpe $\widehat{SR}$ has, for possibly non-normal returns, an asymptotic standard error

$$\hat\sigma(\widehat{SR}) = \sqrt{\frac{1 - \hat\gamma_3\,\widehat{SR} + \frac{\hat\gamma_4 - 1}{4}\,\widehat{SR}^{\,2}}{n-1}},$$

where $n$ is the number of return observations, $\hat\gamma_3$ is the skewness of returns, and $\hat\gamma_4$ is the kurtosis (not excess; $\hat\gamma_4 = 3$ for a normal). Negative skew and fat tails, the norm for real strategies, raise the standard error, making a given Sharpe less trustworthy. The Probabilistic Sharpe Ratio is the probability that the true Sharpe exceeds a benchmark $SR^\star$:

$$\widehat{PSR}(SR^\star) = \Phi\!\left(\frac{(\widehat{SR} - SR^\star)\sqrt{n-1}}{\sqrt{1 - \hat\gamma_3\,\widehat{SR} + \frac{\hat\gamma_4 - 1}{4}\,\widehat{SR}^{\,2}}}\right),$$

with $\Phi$ the standard-normal CDF. PSR tells you how confident to be that the Sharpe clears a fixed bar, correcting for track-record length and non-normality.

Step 2: the benchmark that accounts for selection

The DSR's key move is to set the benchmark $SR^\star$ not to zero but to the expected maximum Sharpe under the null that no strategy has skill, given that you tried $N$ of them. Across $N$ trials with per-trial Sharpe variance $V\!\left[\{\widehat{SR}_n\}\right]$, extreme-value theory (the maximum of $N$ Gaussians follows a Gumbel law) gives the expected maximum:

$$SR^\star_0 = \sqrt{V\!\left[\{\widehat{SR}_n\}\right]}\;\Big[(1-\gamma)\,\Phi^{-1}\!\Big(1 - \tfrac{1}{N}\Big) + \gamma\,\Phi^{-1}\!\Big(1 - \tfrac{1}{N}e^{-1}\Big)\Big],$$

where $\gamma \approx 0.5772$ is the Euler–Mascheroni constant and $\Phi^{-1}$ is the inverse standard-normal CDF. The bracket is the expected maximum of $N$ standard normals; multiplying by the cross-trial Sharpe standard deviation $\sqrt{V[\{\widehat{SR}_n\}]}$ scales it to your actual dispersion of results. This benchmark rises with $N$: the more variants you tried, the higher a Sharpe you should expect from luck alone, and the higher the bar your winner must clear.

Step 3: deflate

The Deflated Sharpe Ratio is just the PSR evaluated at this selection-aware benchmark:

$$\boxed{\;DSR = \widehat{PSR}(SR^\star_0) = \Phi\!\left(\frac{(\widehat{SR} - SR^\star_0)\sqrt{n-1}}{\sqrt{1 - \hat\gamma_3\,\widehat{SR} + \frac{\hat\gamma_4 - 1}{4}\,\widehat{SR}^{\,2}}}\right).\;}$$

Read it as the probability that the strategy's true Sharpe is positive after accounting for the number of trials, the non-normality of returns, and the sample length. A DSR near 1 is evidence of genuine skill; a DSR near 0.5 or below means the observed Sharpe is indistinguishable from the best of $N$ lucky draws. A conventional acceptance threshold is $DSR > 0.95$.

Worked example

A backtest reports a per-annum Sharpe of $\widehat{SR} = 1.5$ over $n = 1{,}250$ daily observations (≈5 years), with mild negative skew $\hat\gamma_3 = -0.5$ and fat tails $\hat\gamma_4 = 6$. You tried $N = 100$ variants whose Sharpes had a cross-trial standard deviation of $0.5$.

First the benchmark. With $N=100$: $\Phi^{-1}(1 - 1/100) = \Phi^{-1}(0.99) \approx 2.326$ and $\Phi^{-1}(1 - e^{-1}/100) = \Phi^{-1}(0.99632) \approx 2.681$. So

$$SR^\star_0 = 0.5\big[(1-0.5772)(2.326) + 0.5772(2.681)\big] \approx 0.5\,[0.983 + 1.548] = 0.5 \times 2.531 \approx 1.27.$$

The expected lucky maximum Sharpe is already $1.27$, most of your observed $1.5$. Plugging into the DSR (using per-observation Sharpe consistently), the numerator $(\widehat{SR} - SR^\star_0)$ is small and the fat tails inflate the denominator, so the DSR lands well below the $0.95$ bar. The verdict: a $1.5$ Sharpe, selected from 100 tries on a 5-year, fat-tailed sample, is not convincing evidence of skill. Had you tried only $N=1$, the benchmark would be $\approx 0$ and the same $1.5$ Sharpe would be strongly significant. The number you must disclose is $N$.

Failure modes

• Undercounting $N$. The trials include every variant you, your team, and the literature ever ran on this data, not just the grid in your final notebook. Undercount $N$ and the deflation is too weak.
• Ignoring skew and kurtosis. Assuming normality understates the Sharpe standard error; real strategies with negative skew and fat tails deserve a wider confidence interval.
• Correlated trials. The variance term $V[\{\widehat{SR}_n\}]$ and the effective independent $N$ are entangled when variants overlap; use the observed dispersion of trial Sharpes rather than a nominal count.
• Short samples. Small $n$ makes $\widehat{SR}$ itself noisy; the $\sqrt{n-1}$ term punishes short track records appropriately.

In interviews

Expect "you have a backtest with a Sharpe of 2, what else do you need to know to judge it?" The three answers the DSR formalizes are: how many strategies you tried ($N$), how long and how non-normal the track record is ($n$, skew, kurtosis), and how dispersed the trial Sharpes were. Be able to explain the expected-maximum benchmark, that the max of $N$ Gaussians grows like $\sqrt{2\ln N}$ (or the exact Gumbel expression above), and why fat tails and negative skew lower your confidence in a given Sharpe. The one-sentence summary: the DSR asks whether your best Sharpe beats the best Sharpe luck would have handed you across all your trials. See Data-Snooping Bias and Backtest Overfitting.