| Sharpe ratio | Deflated Sharpe ratio | |
|---|---|---|
| What it is | A ratio of excess return to volatility | A probability that the observed result is not luck |
| Range | Any real number | 0 to 1 |
| Counts how many variants were tried | No | Yes — this is the main correction |
| Accounts for sample length | No | Yes |
| Accounts for skew and fat tails | No | Yes |
| Can be raised by testing more ideas | Yes, mechanically | No — more trials push it down |
| Comparable across strategies | Only if the search behind each was identical | Yes, by construction |
How is it different from the Sharpe ratio?
An ordinary Sharpe ratio describes a track record. It says nothing about how that track record was found. Two strategies can post the same Sharpe when one was the only idea its author ever tested and the other was the best of fifty thousand — and only one of those numbers is evidence.
The deflated Sharpe ratio closes that gap. It asks what the highest Sharpe you would expect to see is, given that many variants were tried and only the winner is being shown, and then asks how likely it is that the observed result beats that expectation. The statistic was introduced by David Bailey and Marcos López de Prado, who also corrected for two properties of real return series that the standard Sharpe assumes away: skew and excess kurtosis.
The practical consequence is that deflating cannot be gamed by working harder. Testing more ideas raises your best raw Sharpe and lowers your deflated one.
What does a deflated Sharpe of 0.2 mean?
It means that after accounting for the search, roughly a one-in-five chance remains that the strategy genuinely beats the benchmark, and roughly four in five that you are looking at the best draw from a pile of noise. It does not mean the strategy has a Sharpe ratio of 0.2. This is the single most common misreading of the statistic, and it runs in the flattering direction: 0.2 sounds like a modest but real result, when it is closer to a coin flip weighted against you.
A strategy can show an in-sample Sharpe of 2.4 and a deflated Sharpe of 0.2 at the same time. Both numbers are correct. The first describes the chart; the second describes how much the chart is worth.
What is a good deflated Sharpe ratio?
There is no universal threshold, and any source that gives you one has stopped being careful. Because the statistic is a probability, the honest question is what confidence you require before acting, and that is set by the decision rather than the strategy.
Three things move the bar. How costly a false positive is for you. How many strategies you intend to run at once, since running twenty candidates at 0.6 is a different exposure from running one. And whether the trial count fed into the calculation was honest — a deflated Sharpe computed from an undercounted search is inflated in exactly the way the statistic exists to prevent.
The convention in the literature is to work at conventional statistical confidence levels, which puts a meaningful result well above 0.9 rather than above 0.5. Treat any value below that as an argument for more evidence, not as a mild endorsement.
What does it need in order to be computed?
Four inputs, and one of them is where the difficulty lives.
- The observed Sharpe ratio of the candidate.
- The length of the return series.
- The skew and kurtosis of those returns.
- The number of independent trials the candidate was selected from — and the variance of the Sharpe ratios across those trials.
The first three are properties of a file. The fourth is a property of a process, and it is why the statistic is quoted far more often than it is computed correctly. If the trial count is guessed, remembered, or taken as the number of strategies you happened to save, the deflation is understated and the resulting number is more comforting than it should be.
What does the deflated Sharpe ratio not tell you?
It does not tell you a strategy will make money, and it does not forecast anything. It is a statement about evidence, not about the future. A high value means the backtest is not easily explained by selection; it does not mean the effect will persist, that the regime will hold, or that costs will not remove it.
It is also blind to look-ahead bias and survivorship bias. Those corrupt the return series itself, and no statistic computed from a corrupted series can detect that it was corrupted. A clean deflated Sharpe on dirty data is a clean measurement of the wrong thing.
Common questions
Is a deflated Sharpe ratio of 0.5 good?
No. Because the statistic is a probability rather than a ratio, 0.5 means the evidence is consistent with a coin flip: about as likely to be selection luck as a real edge. Intuition from ordinary Sharpe ratios, where 0.5 is a modest positive, points the wrong way here.
Who invented the deflated Sharpe ratio?
It was introduced by David Bailey and Marcos Lopez de Prado, as part of a body of work on backtest overfitting that also produced the probability of backtest overfitting and the minimum track record length.
Can I compute a deflated Sharpe ratio for a strategy someone else built?
Only if they tell you how many variants they tested, and most will not, because most did not count. Without a trial count the calculation can still be run, but it silently assumes a search far smaller than the real one and returns a number that is too generous.
How does it relate to the probability of backtest overfitting?
They answer adjacent questions from opposite directions. The deflated Sharpe ratio asks how likely this particular result is to be real given the search; the probability of backtest overfitting asks how often, across repeated splits, the in-sample winner turns out to be below median out-of-sample. The second is a property of the whole search rather than of one candidate.