Probabilistic Sharpe Ratio (PSR)
The probability that a portfolio's true Sharpe ratio exceeds a chosen benchmark, given the Sharpe you actually observed. It replaces a single number with a confidence statement, correcting for the four things that make a raw Sharpe misleading: a short sample, skewed returns, fat tails, and serially correlated returns.
Introduced by Bailey and López de Prado (2012). PSR is the foundation the Deflated Sharpe Ratio (DSR) is built on, and it shares its standard error with the Sharpe Stability Ratio (SSR).
The Problem It Solves
A reported Sharpe ratio is an estimate, computed from a finite sample of returns. Two portfolios can report the same Sharpe of 1.2 while one has three decades of history and the other has eight months. The point estimate cannot tell them apart; the sampling distribution can.
The textbook standard error of the Sharpe estimator assumes returns are independent and Normally distributed. Financial returns are neither. Negative skewness and excess kurtosis both inflate the true uncertainty of the estimate, which means the naive standard error understates how much of your Sharpe could be luck. Positive serial correlation does the same thing more aggressively: it reduces the number of genuinely independent observations you have, so a month-count overstates your effective sample size.
PSR takes the observed Sharpe, builds the sampling distribution that accounts for all four effects, and reports a single probability: how likely is it that the true Sharpe is above the benchmark? A high Sharpe on a short, lumpy, autocorrelated track record produces a low PSR, which is the honest answer.
Mathematical Formulation
The PSR itself
Let be the observed Sharpe ratio, the benchmark Sharpe you want to beat (usually zero), and the standard error of the estimator. Then
where is the standard-Normal CDF. The whole content of the method sits in the standard error, which is where the higher moments and the autocorrelation enter.
The generalized variance of the estimator
FolioLab uses the autocorrelation-aware form of the Sharpe estimator variance. With observations, skewness , kurtosis (not excess kurtosis: a Normal distribution gives 3), and first-order autocorrelation :
Read the three terms in order. The first is the variance inflation from serial correlation alone: at it equals one, and it grows without bound as . The second penalizes negative skewness (a negative makes the term positive, so left-tailed return distributions raise the standard error). The third penalizes fat tails, and it scales with , so the higher the Sharpe the more a heavy tail costs you.
Setting collapses this to the familiar independent-observations form quoted in most of the literature:
A Normal, serially uncorrelated series (, , ) reduces it further to , the classical Lo (2002) result.
Minimum Track Record Length (MinTRL)
The same machinery inverts into a sample-size question. If the observed Sharpe persists, how long must the track record be before PSR clears ? Writing for the bracketed multiplier above:
FolioLab uses and reports the answer both in observations and in years. When the requirement is infinite: no amount of additional history rescues a Sharpe that does not beat the benchmark in the first place.
Confidence interval
The two-sided interval for the Sharpe estimator uses the same standard error:
An interval whose lower bound sits below zero is the same message as a PSR below 0.5, stated in Sharpe units rather than probability units.
What FolioLab Computes
Every optimization run builds a Sharpe inference payload from the portfolio return series and renders it as the PSR card on the results page. Alongside the probability itself the payload carries the inputs, so the number is auditable rather than a black box: the periodic and annualized Sharpe, the standard error, skewness, kurtosis, first-order autocorrelation, the observation count, MinTRL in observations and years, and the confidence-interval bounds.
Two scales are in play, deliberately. The estimator works on the periodic (per-observation) Sharpe scale, because that is the scale on which counts. Benchmarks, confidence bounds and the reported Sharpe are all on the annualized scale, converted by for daily data. A benchmark you supply as an annualized Sharpe is divided back down internally before the comparison happens.
The guards matter for reading edge cases. Kurtosis is floored at 1 and is clamped to , since the variance formula divides by and a near-unit autocorrelation would otherwise blow the standard error up to infinity. If the computed variance underflows, PSR degenerates to a hard 1 or 0 depending on which side of the benchmark the Sharpe falls. A series shorter than two observations, or with zero volatility, produces no payload at all rather than a fabricated one.
One display caveat: when a stored run predates the backend inference payload, the results page computes PSR client-side with and marks the card with a footnote. That is the conservative direction for a positively autocorrelated series, but it is an approximation, and the card says so.
How To Read It
PSR is a probability in . Higher is better. Against the default benchmark of zero it answers: how confident can I be that this portfolio has any risk-adjusted edge at all?
| PSR | Verdict | Interpretation |
|---|---|---|
| Statistically credible | The track record is long enough, and the return distribution well behaved enough, to support the Sharpe at a conventional confidence level. | |
| Suggestive, not conclusive | The Sharpe is more likely real than not, but sampling noise remains a live explanation. Read MinTRL to see how much more history would settle it. | |
| Not supported | The observed Sharpe is at or below the benchmark once the estimator's uncertainty is priced in. Treat the headline number as noise. |
Read PSR beside MinTRL, always. A PSR of 0.72 with a MinTRL of 400 observations on a 250-observation sample tells you the shortfall is roughly another year of data. A PSR of 0.72 with a MinTRL of 9,000 observations tells you the effect is too small relative to its noise to ever be settled by waiting.
PSR, DSR and SSR
Three Sharpe diagnostics, one shared standard error, three different questions. None of them substitutes for another.
| Metric | Question | Benchmark |
|---|---|---|
| PSR | Is this Sharpe credible at this sample size and shape? | , or a benchmark you choose |
| DSR | Did I cherry-pick it from N methods? | , the expected maximum under zero skill |
| SSR | Is it stable across time, or driven by one episode? | HAC-robust dispersion of the rolling Sharpe |
The cleanest way to hold the relationship: DSR is PSR with a raised bar. Both evaluate the same expression with the same standard error; DSR simply replaces the zero benchmark with the Sharpe you would expect the best of skill-free trials to produce. When the two coincide exactly. A portfolio can clear PSR comfortably and still fail DSR, which means the track record is long enough but the search that found it was wide enough to explain the result.
Advantages & Limitations
Advantages
- Turns a point estimate into a statement of confidence: a probability is directly comparable across portfolios of different lengths.
- Prices non-Normality explicitly: negative skew and fat tails widen the sampling distribution instead of being ignored.
- Handles serial correlation: smoothed or persistent return series get the effective-sample penalty they deserve.
- Closed form: no bootstrap, no simulation, no tuning parameters beyond the confidence level.
- Comes with an actionable companion: MinTRL converts a failed test into a concrete data requirement.
Limitations
- Asymptotic: the Normal approximation to the estimator's distribution is a large-sample result, and it is least reliable exactly where it matters most, on very short samples.
- Blind to selection bias: PSR has no idea how many methods you tried before keeping this one. That is what DSR is for.
- Only first-order autocorrelation: is a lag-1 estimate. Longer-memory dependence is not captured by this correction.
- Higher moments are themselves estimated: skewness and kurtosis from a short sample are noisy, and that noise is not propagated into the PSR.
- In-sample: it validates the statistical strength of a historical record. It is not a forecast, and it says nothing about whether the process that generated those returns still holds.
References
- Bailey, D. H., & López de Prado, M. (2012). "The Sharpe Ratio Efficient Frontier." Journal of Risk, 15(2), 3-44. doi:10.21314/JOR.2012.255.
- Lo, A. W. (2002). "The Statistics of Sharpe Ratios." Financial Analysts Journal, 58(4), 36-52. doi:10.2469/faj.v58.n4.2453.
- Mertens, E. (2002). "Comments on the Variance of the IID Estimator in Lo (2002)." Working Paper.
- Opdyke, J. D. (2007). "Comparing Sharpe Ratios: So Where are the p-values?" Journal of Asset Management, 8(5), 308-336. doi:10.1057/palgrave.jam.2250084.
Not investment advice. Past performance is not indicative of future results.