The Deflated Sharpe ratio (and PSR), explained
A backtested Sharpe ratio is almost always too good. The Deflated Sharpe ratio asks the honest question: after you tried dozens of variations and kept the best one, how much of that Sharpe was skill — and how much was the luck you were bound to find eventually?
Why a backtested Sharpe lies
The Sharpe ratio of the best configuration you tested is upward-biased, and the more variations you try, the worse the bias gets. If you test one strategy, a high Sharpe means something. If you test two hundred and report the winner, the winner is partly the product of selection bias — the maximum of many noisy numbers drifts up even when none of the strategies has a real edge. This is the multiple-testing problem: more trials → a higher “lucky” Sharpe you would expect from pure chance. Every parameter you sweep and every variation you breed quietly raises the bar the real edge has to clear.
The Probabilistic Sharpe Ratio (PSR)
Before deflating for trials, you need a way to say how confident you are in a single Sharpe. That's the Probabilistic Sharpe Ratio: the probability that the strategy's true Sharpe exceeds a chosen benchmark, given three things about the track record.
- Length — a Sharpe measured over ten trades is far shakier than the same Sharpe over a thousand. Longer samples earn more confidence.
- Skew — negatively skewed returns (many small gains, rare large losses) make a high Sharpe less trustworthy.
- Kurtosis — fat tails mean the extremes happen more often than a bell curve implies, so the Sharpe understates the real risk.
PSR turns those into a single probability in [0, 1]. A PSR of 0.95 against a zero benchmark reads “95% confident the true Sharpe is above zero.”
The Deflated Sharpe Ratio (DSR)
The Deflated Sharpe ratio is PSR with the benchmark set to the expected maximum Sharpe you'd get from your search under the null of no edge. Instead of asking “is the Sharpe above zero,” it asks “is the Sharpe above the best you'd expect to find by luck after this many trials.” It discounts your Sharpe for four things at once:
- The number of configurations tried (N) — more trials, higher expected lucky maximum.
- The variance across those trials' Sharpes — a wide spread of results means the maximum was easier to inflate.
- Skew and kurtosis — the same fat-tail and asymmetry penalties PSR applies.
- Sample length — short track records get less benefit of the doubt.
The output is a probability that the strategy's true Sharpe is above zero after accounting for selection. The method is due to Bailey and López de Prado (“The Deflated Sharpe Ratio,” 2014).
DSR = PSR( observed Sharpe ; benchmark = E[ max Sharpe over N trials ] )
inputs → N (trials) · Var(Sharpe across trials) · skew · kurtosis · sample length
How to read it
The Deflated Sharpe ratio is a probability, not a ratio like the raw Sharpe. Read it the way you'd read a confidence level.
- Near or above ~0.95 — strong evidence the edge is real and survived the discount for how hard you searched.
- Around 0.5 or below — your Sharpe is plausibly just the best of many lucky draws; treat the strategy as unproven.
The practical lesson writes itself: every parameter you optimize and every variation you breed raises N, so keep the search small, and always deflate before you believe a Sharpe.
The deflated-Sharpe checklist
- Did you count every configuration you tried — optimization and breeding included?
- Is the Sharpe still positive after the multiple-testing haircut?
- Are the returns fat-tailed or negatively skewed?
- Is the sample long enough to earn confidence?
- Did it also survive Walk-Forward and Monte-Carlo?
How QUANTHEON Lab does this for you
The Overfitting Polygraph applies a Deflated-Sharpe haircut that counts the configurations you actually tried in a parameter search — and folds it together with Walk-Forward and Monte-Carlo into a single verdict: trustworthy, fragile or curve-fit. Every backtest also carries a reproducibility receipt, so the number you saw is the number anyone can reproduce. The deflation is instant; the full verdict needs all three checks run.
Bred variants are corrected too, but not by this number — and the distinction is worth knowing, because pooling them would be the wrong statistic. A parameter search produces one family of configurations of a single idea, so the right correction is deflating that idea's best Sharpe for how many settings were tried. Evolving a population produces many different strategies, so the right correction is a family-wise one: every survivor is judged once against a sealed out-of-sample vault it has never touched, and the search itself is taxed. Two searches, two corrections — the one thing neither is allowed to be is absent.
FAQ
What is the Deflated Sharpe ratio?
It's the Probabilistic Sharpe Ratio measured against the maximum Sharpe you'd expect to find by luck after N trials. It returns the probability that a strategy's true Sharpe is above zero once you account for how many configurations were tried, their spread, the returns' skew and kurtosis, and the sample length.
What's the difference between PSR and DSR?
PSR is the probability the true Sharpe beats a fixed benchmark, given track-record length, skew and kurtosis. DSR is PSR with that benchmark raised to the expected maximum Sharpe under N trials — so it additionally penalizes multiple testing. DSR is PSR corrected for how hard you searched.
What is a good Deflated Sharpe ratio?
Because it's a probability in [0,1], a value near or above 0.95 is strong evidence the edge is real. Around 0.5 or below means the Sharpe is plausibly just luck from many trials, and the strategy should be treated as unproven.
Related: What is overfitting? · Walk-Forward analysis · Monte-Carlo simulation · Sharpe ratio calculator