Definition
The deflated Sharpe ratio adjusts the interpretation of an observed Sharpe ratio for non-normal returns and the selection effect created by trying multiple investment variants. It compares the observed statistic with a benchmark that reflects skewness, kurtosis, sample length and an effective number of trials. The exact finite-sample convention must be fixed before use.
How it is calculated
Deflated Sharpe ratio = standard normal CDF Phi of (observed Sharpe S-hat minus S-star) divided by the standard error of the observed Sharpe. Under zero-mean approximately normal independent trial Sharpes, S-star = v times [(1-c) Phi-inverse of (1-1/N) + c Phi-inverse of (1-1/(N e))], where v is trial-Sharpe standard deviation, c is approximately 0.577216, N is effective trials greater than 1, and e is Euler's number. For one zero-mean trial, use benchmark zero and the G48 asymptotic standard-error convention.
| Symbol | Meaning | Unit |
|---|---|---|
| DSR | deflated Sharpe probability score | probability |
| Phi | standard normal cumulative distribution function | probability |
| S | observed Sharpe ratio | dimensionless |
| S^* | trial-adjusted benchmark Sharpe ratio | dimensionless |
| se | standard error | Sharpe-ratio units |
| v | trial-Sharpe standard deviation | dimensionless |
| c | Euler-Mascheroni constant, approximately 0.577216 | dimensionless |
| N | effective number of trials, greater than 1 | count |
-
Estimate the observed Sharpe ratio and its sampling uncertainty.
-
Set the trial-adjusted benchmark using the stated effective-trial convention.
-
Standardize the gap from that benchmark and apply the standard normal cumulative distribution function.
Interpretation
A high Sharpe found after searching many variants is less surprising than the same Sharpe specified in advance. Deflation moves the comparison point to account for that search and for non-normal return shapes. It cannot infer the search history from the Sharpe alone.
Assumptions and conventions
-
The Sharpe convention, return frequency and finite-sample standard-error approximation are specified.
-
The trial count or effective-trial model represents how the selected result was produced.
-
The return distribution's skewness and kurtosis are estimated consistently with the Sharpe sample.
The Sharpe ratio and its benchmark are dimensionless; DSR is a probability.
Limitations and common misreadings
-
There is no universal trial-count cutoff.
-
DSR does not establish that a selected strategy will generalize.
References
- David H. Bailey, Marcos López de Prado. (2014). The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality. The Journal of Portfolio Management. Whole article. doi:10.3905/jpm.2014.40.5.094. Source

