Deflated Sharpe ratio
The deflated Sharpe ratio corrects an observed Sharpe ratio for the number of configurations that were tried before it was selected, the length of the sample and the shape of the return distribution. It expresses the probability that the observed Sharpe exceeds what selection alone would have produced on noise.
Senzoukria · Glossary · Updated September 2026
The problem it addresses
Run enough configurations on the same data and the best one will show a positive Sharpe ratio even when every configuration is random. The expected maximum of many random Sharpe ratios grows with the number of trials. Keeping the winner and reporting its Sharpe therefore overstates the evidence.
The deflated Sharpe ratio, introduced by Bailey and López de Prado, computes the Sharpe threshold that selection alone would have produced given the number of trials and the variance across them, then evaluates the observed Sharpe against that threshold with the sample length, skewness and kurtosis of the returns. The output is a probability rather than a ratio.
Inputs and what changes them
- Number of configurations tried: more trials raise the threshold. Sweeping a fourth parameter can turn a solid result into noise without changing the data.
- Number of out-of-sample trades: a longer sample narrows the uncertainty of the observed Sharpe.
- Skewness and kurtosis: fat-tailed or negatively skewed returns widen the uncertainty and lower the probability.
- The probabilistic Sharpe ratio (PSR) is the same calculation against a fixed benchmark instead of the selection-derived threshold.
In Senzoukria
The backtest panel reports a block titled Deflated Sharpe — probability of a real edge after a parameter sweep. Its detail line reads: Sharpe per trade against a threshold produced by selection alone, over the number of configurations and out-of-sample trades used. The verdict text has three levels: solid (the result survives the correction for how many configurations were tried), fragile (the number of configurations tried is enough to explain a good part of it) and noise (not distinguishable from the best of a random draw).
The go/no-go gauntlet of the same panel lists 'Deflated Sharpe: survives the number of trials' as one criterion, with PSR, DSR and the count of configurations tried shown beside it.
Common mistakes
- Under-counting trials. Every configuration you looked at and discarded still counts, including runs made in an earlier session on the same data.
- Reading the DSR as a probability of future profit. It is the probability that the observed Sharpe was not produced by selection, under the stated assumptions.
- Narrowing the sweep after seeing the result to make the DSR improve. That is another trial and should be counted.
- Applying it to a series with too few trades for the moments to be estimated.
Related
- Futures backtesting guide
- Sharpe ratio
- Probability of backtest overfitting
- How to test a trading strategy
This page in other languages
Frequently asked questions
- What is the difference between PSR and DSR?
- The probabilistic Sharpe ratio (PSR) gives the probability that the true Sharpe exceeds a benchmark you choose, accounting for sample length, skewness and kurtosis. The deflated Sharpe ratio (DSR) is a PSR whose benchmark is derived from the number of trials and their variance, so it also corrects for selection. The desktop gauntlet shows both values.
- Why does a longer sweep lower my deflated Sharpe?
- Because the threshold produced by selection alone rises with the number of configurations. A Sharpe that stood out among ten trials is ordinary among four hundred. The observed Sharpe did not change; the evidence it represents did. Reducing the number of swept parameters or lengthening the out-of-sample period are the two ways to strengthen the result.