What is the Deflated Sharpe Ratio?
The Deflated Sharpe Ratio (DSR) is the Probabilistic Sharpe Ratio with the benchmark raised to the Sharpe you would expect to see by luck alone after testing many strategies. If you try enough variations, the best one will look good whether or not it has any edge. DSR asks whether your winner beats that luck threshold. A strategy with PSR near 1.00 against a zero benchmark can fall below 0.30 once 1,000 trials are accounted for.
What problem does it solve?
The Probabilistic Sharpe Ratio corrects for track-record length and return shape, but it evaluates one track record in isolation. That is the wrong question if the strategy in front of you is the survivor of a search.
Suppose you test 1,000 parameter combinations on the same data. Even if none has genuine edge, the sample Sharpe ratios will scatter around zero, and the best of the 1,000 will be well above zero purely from sampling variation. Reporting that winner as though it were the only thing you tried is selection bias, and it is the single most common way backtests mislead.
Statisticians call this the multiple testing problem, and it is the same issue that produces false positives in clinical trials and particle physics. Finance arrived at it late and still under-corrects for it.
Bailey and López de Prado's fix is direct: instead of testing against zero, test against the Sharpe you would expect the best of N trials to produce under the null hypothesis of no skill.
Formula
DSR is PSR evaluated at a deflated benchmark:
Two inputs deserve attention because they are where the method is easiest
to abuse. N is the number of trials you really ran,
including the ones you abandoned early and the ones you would rather forget;
understating it inflates DSR. sd(SR) is the dispersion of Sharpe
ratios across those trials, which requires you to have kept them.
Worked example
A strategy with an observed Sharpe of 1.50 over 120 monthly observations, skewness −0.8 and kurtosis 6.0. Sharpe ratios across the trials had a standard deviation of 0.50. The denominator term is the same as for PSR:
Now watch what the trial count does to the same strategy:
| Trials (N) | Expected max Sharpe SR*₀ | DSR | Verdict at 0.95 |
|---|---|---|---|
| 1 | — | 1.0000 | Passes (no selection to correct) |
| 10 | 0.7873 | 0.9997 | Passes |
| 100 | 1.2653 | 0.8736 | Fails |
| 1000 | 1.6276 | 0.2671 | Fails badly |
Nothing about the strategy changed between those rows. Only the honesty of the accounting did. Against a zero benchmark this track record gives a PSR of essentially 1.0000 — which is why a naive PSR on a mined strategy is close to meaningless.
The N = 1 row is a genuine degenerate case: with a single trial there is no selection bias, the expected maximum is undefined, and DSR collapses back to ordinary PSR. That is correct behaviour, not a gap.
PSR vs DSR: which do I need?
| PSR | DSR | |
|---|---|---|
| Question answered | Is this Sharpe distinguishable from the benchmark? | Is this Sharpe distinguishable from the best of N lucky tries? |
| Benchmark | Chosen by you (often 0) | Expected maximum under the null |
| Corrects for track length | Yes | Yes |
| Corrects for skew/kurtosis | Yes | Yes |
| Corrects for selection bias | No | Yes |
| Needs a trial count | No | Yes — and an honest one |
| Use when | You have one strategy, designed in advance | You searched, optimised or selected |
The practical rule: if you can honestly say you tested exactly one specification and are reporting it whatever the result, PSR is enough. The moment you swept a parameter, PSR overstates your evidence.
Limitations
- N is self-reported and unverifiable. The correction is only as honest as the trial count you feed it. There is no way for a reader to audit it, which is why the number should be disclosed alongside the result rather than folded silently into a score.
- Trials are assumed independent. Sweeping one parameter produces highly correlated strategies, so the effective number of independent trials is smaller than the raw count. Using the raw count is conservative; treating correlated variants as independent overstates the penalty.
- It inherits every PSR assumption. IID returns, known higher moments, and no account of costs, capacity or regime change.
- It does not rescue a bad process. DSR quantifies the damage from searching; it does not undo it. The stronger discipline is to pre-register the specification.
References
- Bailey, D. & López de Prado, M. (2014). “The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality.” Journal of Portfolio Management 40(5), 94–107.
- Bailey, D. & López de Prado, M. (2012). “The Sharpe Ratio Efficient Frontier.” Journal of Risk 15(2), 3–44.
- Bailey, D., Borwein, J., López de Prado, M. & Zhu, Q. (2014). “Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance.” Notices of the AMS 61(5), 458–471.
- Harvey, C. & Liu, Y. (2015). “Backtesting.” Journal of Portfolio Management 42(1), 13–28.