Short answer
The deflated Sharpe ratio (DSR), from Bailey and Lopez de Prado (2014), adjusts an observed Sharpe ratio for how many strategy variants were tried, the track length, and non-normal returns (skew and kurtosis). It outputs a probability that the true Sharpe beats what the best trial would reach by luck, typically read against a 95% threshold.
This article is educational and general. It explains a statistical technique for evaluating track records. It is not personalised investment advice, not a recommendation to buy, sell, or hold any security, and not a performance claim about any specific strategy.
Search hard enough and a beautiful Sharpe ratio appears from pure chance, like a face in the clouds. The number is real; the skill it seems to promise is not. The deflated Sharpe ratio is the arithmetic that tells the two apart.
The short version
The deflated Sharpe ratio (DSR) adjusts an observed Sharpe ratio for three things a raw figure hides: how many strategy variants were tried, how long the track is, and how non-normal the returns are. It converts that into a single probability that the true Sharpe beats the higher level the best of all those trials would be expected to reach by luck alone. A high raw Sharpe with a low deflated confidence is a warning, not a result: it means the search, not the skill, most likely produced the number.
Why the raw Sharpe overstates skill after a search
Assume you have zero skill. You generate a hundred random trading rules, backtest them all, and keep the best one. The winner will have a flattering Sharpe ratio: you took the maximum of a hundred draws of noise. The expected maximum of many random Sharpe ratios sits well above zero, and it climbs as you test more variants. This is the same engine that drives the multiple-testing problem in general: the more hypotheses you check, the higher the best one scores by luck alone, so the winner’s headline number is a biased estimate of its true quality. This is the mechanism behind backtest overfitting, where a search tunes a rule to fit the historical noise rather than any durable signal, and the flattering in-sample Sharpe does not survive contact with new data.
Assuming Sharpe is a new idea to you would be a mistake here, so we will not re-derive it; if you want the definition and how to read a 1 versus a 3, start with the Sharpe explainer. The point of this piece is narrower. A Sharpe ratio, taken alone, answers “how efficient was this curve?” It cannot answer “how many curves did you look at before you showed me this one?” And that second question is the one that decides whether the first answer is trustworthy.
What the deflated Sharpe ratio corrects for
David Bailey and Marcos López de Prado formalised the fix in 2014. The deflated Sharpe ratio takes the observed Sharpe and shrinks the bar it must clear using three inputs:
- The number of independent trials. The more distinct strategy configurations that were tried in the search, the higher the Sharpe you would expect to see from luck, and so the higher the hurdle the real candidate must clear.
- The length of the track. A short sample gives a noisy Sharpe estimate. The more observations behind the number, the more precise it is, and the less the deflation has to punish it.
- The shape of the returns: skew and kurtosis. The standard error of a Sharpe estimate is inflated by negative skew and by fat tails. A strategy whose returns are lumpy and crash-prone, exactly the profile of a concentrated equity book, has a wider error band around its Sharpe than the textbook normal case assumes, so a fat-tailed curve needs a bigger raw Sharpe to reach the same confidence.
Feed those in and the DSR gives you a probability: the chance that the true Sharpe beats the level the best of your trials would hit by luck alone, given everything you tried and everything the return distribution is doing. That benchmark is the expected-maximum-of-noise hurdle, which itself sits above zero and rises with the trial count. The output is expressed as a confidence, typically read against a conventional threshold such as 95%.
A worked contrast makes the stakes concrete. Take a raw Sharpe of 3 earned over ten years from just two candidate variants: the sample is long, the search was shallow, and deflation barely touches it, so the confidence stays high. Now take the very same raw Sharpe of 3 earned over two years from two hundred variants: the sample is short and the expected-maximum hurdle from two hundred trials is large, so the same headline number can deflate to a coin-flip confidence. Identical raw Sharpe, opposite verdicts, because the DSR is pricing in what you did not show: the length of the track and the size of the search behind it.
How to read it: a confidence, not a score
This is the part people miss. The deflated Sharpe ratio is not a smaller, more honest Sharpe number that you then compare against the usual benchmarks. It is a probability that the edge is real at all. The output answers a yes-or-no question with a percentage: does the underlying Sharpe clear the bar the search alone would be expected to reach, or did the search manufacture it? Reading the DSR as if it were just a haircut on the performance figure throws away the whole point, which is to separate the magnitude of a result from the credibility of it.
In practice this reframes how a track record should be presented. The Sharpe tells you how smooth the curve was. The deflated Sharpe tells you how much of that smoothness you are entitled to believe, given the way it was found. Both belong next to each other. A performance number shown without any sense of how many attempts produced it is a number chosen for you.
The companion question: how long until a Sharpe is believable
The DSR asks whether a Sharpe is real given the track you have. The obvious next question is the inverse: how long a track do you need before a given Sharpe becomes believable at all? Bailey and López de Prado answer that with the Minimum Track Record Length, the number of observations required for the deflated confidence to cross your chosen threshold. A high Sharpe with a short history can still fail the test simply because there is not yet enough data to rule out luck.
The two tools are two sides of one coin, and we keep them separate on purpose: the deflated Sharpe deflates for how much you searched, and the minimum track record length tells you how long you must wait for the answer to firm up. A track that is too short fails because the evidence has not accumulated yet, and honest evaluation says so rather than rounding the confidence up.
How this fits how we validate our own work
The reason this technique matters to us is simple: we have searched. The research log is a record of many configurations tried, and pretending otherwise would be self-deception. So we never report a Sharpe ratio as if it stood alone. It sits next to the multiple-testing-adjusted significance, because a figure that ignores the number of attempts behind it is closer to advertising than to evidence. The way we frame the underlying significance test, and the distribution-free bootstrap we lean on because our returns are fat-tailed rather than normal, is laid out in is the edge real, or just luck? The deflated Sharpe is the Sharpe-specific member of that same family: adjust for trials, adjust for sample, adjust for shape, then state a confidence rather than a boast.
And because an in-sample confidence, however carefully deflated, is still an in-sample number, the test that actually settles the question is the forward one. Our live signals and net asset value are published as they happen, so the record accrues out of sample where no search can touch it. You can read the running track on the track record page and the timestamped commitments on the verification log.
The honest limit: the trial count is itself an estimate
The deflated Sharpe ratio has a soft spot at its centre, and we would rather name it than hide it. The most important input, the number of independent trials, is genuinely hard to know. Real strategy searches are a tangle of overlapping variants: nudge one parameter and the new backtest is heavily correlated with the old one, so it does not count as a full extra trial, but nor is it free. Estimating the effective number of independent trials from a correlated search is an art with real error bars, and the deflation is only as good as that estimate. Undercount the trials and you flatter the result; overcount them and you bury a real edge.
So the deflated Sharpe ratio is a discipline, not a guarantee. It forces the right question, how much did the search inflate this number, and it gives a defensible answer, but the answer inherits the uncertainty of the trial count it was fed. That is exactly why the deflated figure earns its place beside the raw Sharpe and the forward, out-of-sample track rather than replacing either. It is one honest estimate among several, and the point of showing all of them is that no single one gets to be the whole story.
Sources & further reading
- Bailey, D. H. & López de Prado, M. (2014). “The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality.” Journal of Portfolio Management, 40(5), 94 to 107.
- Bailey, D. H. & López de Prado, M. (2012). “The Sharpe Ratio Efficient Frontier.” Journal of Risk, 15(2), 3 to 44. (Introduces the Minimum Track Record Length.)
- On the multiple-testing hurdle that motivates deflating for the trial count: Harvey, C. R., Liu, Y. & Zhu, H. (2016). “... and the Cross-Section of Expected Returns” (the ellipsis is the paper’s own stylised title). Review of Financial Studies, 29(1), 5 to 68.