Bailey-López de Prado (2014) The Deflated Sharpe Ratio

Summary

As data, machine learning, and computing let analysts backtest millions or billions of strategy configurations, backtest optimizers find parameter combinations that maximize simulated historical performance — producing backtest overfitting. More generally, reporting only positive outcomes (selection bias) without controlling for the number of trials yields over-optimistic performance. Bailey and López de Prado propose the Deflated Sharpe Ratio (DSR), which corrects for two leading sources of performance inflation — selection bias under multiple testing, and non-Normally distributed returns — and so helps separate genuine empirical findings from statistical flukes.

Key Claims

• Two sources of inflation. Reported Sharpe ratios are inflated by (i) selection bias under multiple testing — trying many strategies and keeping the best — and (ii) non-Normal returns (skew/kurtosis), which the ordinary SR ignores. DSR corrects both.
• DSR = PSR against the expected maximum. DSR is the Probabilistic Sharpe Ratio evaluated with the benchmark $SR^*$ set to the expected maximum Sharpe ratio across $N$ trials, estimated by extreme-value theory from $N$ and the cross-sectional variance of the trial Sharpe ratios.
• Minimum Backtest Length (MinBTL). A companion bound: the shortest backtest length for which running $N$ trials would not, by chance alone, be expected to produce a spuriously high in-sample Sharpe ratio.
• Purpose. Turn a raw backtested Sharpe ratio into a multiplicity- and non-Normality-adjusted probability of genuine skill.

Concepts Introduced or Extended

• Probabilistic Sharpe Ratio

Entities Mentioned

• David H. Bailey
• Marcos López de Prado

Quotes

"The Deflated Sharpe Ratio (DSR) corrects for two leading sources of performance inflation: Selection bias under multiple testing and non-Normally distributed returns. In doing so, DSR helps separate legitimate empirical findings from statistical flukes."

My Take

DSR is the operational punchline of the PSR machinery: once you accept that a Sharpe ratio has a standard error, the obvious next question is "best of how many tries?", and DSR answers it by benchmarking against the Sharpe ratio you'd expect to see by luck alone after $N$ trials. That reframing — comparing a discovery not to zero but to the expected maximum of a null search — is exactly the multiple-testing logic behind the factor zoo's haircuts (Harvey-Liu-Zhu), arriving independently from the backtesting side. The practical catch is $N$: the "number of trials" is rarely known and easy to understate (every discarded configuration counts), so DSR is a lower bound on the correction unless the search is honestly logged. Still, it converts a vague warning about overfitting into a number a risk committee can act on.