The Probabilistic Sharpe Ratio

A Sharpe ratio computed from a backtest is an estimate, not a fact. You measured it on one finite sample, and a different sample would have handed you a different number. So the honest question is never "is the Sharpe 1.2?" but "given this Sharpe of 1.2, how sure am I the strategy's true Sharpe is even positive?" The Probabilistic Sharpe Ratio (PSR) answers exactly that: it converts an observed Sharpe into a probability that the real, long-run Sharpe clears some benchmark.

The trick is knowing how noisy your Sharpe estimate is. A Sharpe measured over 30 years is far more trustworthy than the same Sharpe measured over 6 months, and returns with fat tails and negative skew make the estimate noisier still. PSR bakes both effects into a single number between 0 and 1.

The formula

The PSR is the probability that the true Sharpe exceeds a chosen benchmark $SR^\star$:

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

Reading it piece by piece: $\widehat{SR}$ is your observed (per-period) Sharpe; $SR^\star$ is the bar you want to beat (often $0$); $n$ is the number of return observations; $\hat\gamma_3$ is the skewness of the returns and $\hat\gamma_4$ is their kurtosis ($\hat\gamma_4 = 3$ for a normal, higher for fat tails); and $\Phi$ is the standard-normal CDF (the bell curve's running total, which turns a z-score into a probability). The numerator asks how far your Sharpe sits above the bar, scaled up by $\sqrt{n-1}$ so a longer record counts for more. The denominator is the standard error of the Sharpe estimate: negative skew and fat tails make it bigger, shrinking your confidence.

The picture below shows the idea. The bell is the range of true Sharpes consistent with what you measured, centered on $\widehat{SR}$. PSR is the shaded mass to the right of the benchmark $SR^\star$, the chunk of that range where the strategy really does have a positive edge.

Worked example

A strategy shows a monthly Sharpe of $\widehat{SR} = 0.30$ over $n = 36$ months, and you want the probability the true Sharpe is positive, so $SR^\star = 0$.

If returns were normal ($\hat\gamma_3 = 0$, $\hat\gamma_4 = 3$): the denominator is $\sqrt{1 - 0 + \tfrac{3-1}{4}(0.30)^2} = \sqrt{1 + 0.045} = 1.022$. The numerator is $0.30 \times \sqrt{35} = 0.30 \times 5.916 = 1.775$. So the z-score is $1.775 / 1.022 = 1.74$, and $\Phi(1.74) \approx 0.96$. A PSR of 0.96, above the usual 0.95 bar, you're fairly confident the edge is real.

Now make the returns realistic with negative skew $\hat\gamma_3 = -0.5$ and fat tails $\hat\gamma_4 = 6$: the denominator grows to $\sqrt{1 - (-0.5)(0.30) + \tfrac{6-1}{4}(0.30)^2} = \sqrt{1 + 0.15 + 0.1125} = 1.124$. The numerator is unchanged at $1.775$, so the z-score falls to $1.775 / 1.124 = 1.58$, and $\Phi(1.58) \approx 0.94$. The PSR slips to 0.94, now below 0.95. Nothing about the return path changed except its shape, yet the same Sharpe is no longer convincing. That is the entire point: a Sharpe from lumpy, crash-prone returns deserves less trust than one from smooth returns.

Where it fits

PSR judges a single strategy against a fixed bar. But if you tried many strategies and kept the best, that best-of-many Sharpe is inflated by selection, and a fixed bar of $SR^\star = 0$ is too easy. The fix is to raise the benchmark to the Sharpe luck alone would produce across all your trials, which turns PSR into the The Deflated Sharpe Ratio. And the flip side of PSR, "how long a record do I need before a given Sharpe is trustworthy?", is answered by the Minimum Backtest Length.

In interviews

If asked "your backtest Sharpe is 1.5, is that good?", the sophisticated answer is: it depends on how long the sample is and how non-normal the returns are, quantified by the PSR, and on how many variants you tried, quantified by the deflated Sharpe. Be able to name the two ingredients of the standard error, the $\sqrt{n-1}$ that rewards long records and the skew/kurtosis term that punishes fat tails, and to state that PSR is just $\Phi$ of a z-score comparing your Sharpe to a benchmark.