The Deflated Sharpe Ratio

Jovana A.

— 5 min read
Share
The Deflated Sharpe Ratio

The Sharpe ratio answers a simple question: how much return did a strategy generate per unit of volatility? It does not answer a second question that matters just as much: how many strategies did you have to try before you found this one? The Deflated Sharpe Ratio (DSR), introduced by Bailey and López de Prado in 2014, corrects the Sharpe ratio for exactly that — selection bias from testing multiple variants, plus the non-normality that real return distributions almost always exhibit.

The problem DSR is built to solve

Run enough variations of a strategy — different entry days, holding periods, stop-loss thresholds, lookback windows — against any dataset, and one combination will outperform by chance alone, even if none of the variants carries genuine signal. This is formalized as the False Strategy Theorem: even if every tested strategy has a true Sharpe ratio of zero, the highest observed Sharpe ratio among them will typically be positive and statistically significant unless corrected for the number of trials.

The intuition: think of N independent Sharpe ratio estimates, each drawn from a distribution centered at zero (the null hypothesis of no skill). As N grows, the maximum of those N draws drifts further from zero, purely as a property of extreme-value statistics — not because any underlying strategy improved. A backtest result reported without the number of trials behind it is unfalsifiable: there's no way to know whether you're looking at signal or the predictable byproduct of search.

The expected maximum Sharpe ratio under the null

DSR formalizes this by first computing what the best Sharpe ratio among N trials would be expected to look like if none of the trials had any real skill. Given N independent trials with Sharpe ratios assumed to be Normally distributed with variance σ², the expected maximum Sharpe ratio under the null is approximated by:

SR₀ (SharpeRatio) = σ · [(1 − γ)·Φ⁻¹(1 − 1/N) + γ·Φ⁻¹(1 − 1/(Ne))]

where:

  • σ is the standard deviation of the Sharpe ratios across the N trials (how much the trials disagree with each other)
  • γ is the Euler-Mascheroni constant (≈ 0.5772)
  • Φ⁻¹ is the inverse standard Normal CDF (the quantile function)
  • e is Euler's number
  • N is the number of independent trials

Two things fall directly out of this formula and are worth internalizing on their own:

  • SR₀ grows with N. More trials raise the bar a strategy must clear to be considered genuine. This is the entire point — a Sharpe ratio of 2.0 selected from 5 trials is a different claim than the same number selected from 500.
  • SR₀ grows with σ. More disagreement between trial outcomes also raises the bar, because high cross-trial variance is itself consistent with a process driven by noise rather than a stable, persistent edge.

From SR₀ to the Deflated Sharpe Ratio

Once you have SR₀ — the expected best-case Sharpe ratio under pure luck — the Deflated Sharpe Ratio asks a probabilistic question: given the sample length and the higher moments (skewness, kurtosis) of the strategy's actual returns, what's the probability that the true Sharpe ratio exceeds SR₀, rather than exceeding zero?

This is the Probabilistic Sharpe Ratio (PSR) framework, evaluated against SR₀ instead of zero:

DSR = Φ ( (ŜR − SR₀) · √(T − 1) / √(1 − γ₃·ŜR + ((γ₄ − 1)/4)·ŜR²) )

where:

  • ŜR is the strategy's observed (non-annualized) Sharpe ratio
  • SR₀ is the expected maximum Sharpe ratio computed above
  • T is the number of return observations in the backtest (the sample length)
  • γ₃ is the skewness of the strategy's returns
  • γ₄ is the kurtosis of the strategy's returns
  • Φ is the standard Normal CDF

The numerator is the gap between what you observed and what luck alone would predict as the best case. The denominator adjusts for sample length and for the fact that Sharpe ratio estimates from skewed, fat-tailed return distributions are noisier than the Normal-distribution assumption underlying the basic Sharpe ratio. Negative skew and excess kurtosis — both common in option-selling and short-volatility strategies — widen this denominator and push DSR down, independent of N.

Why this denominator matters as much as the trial count

It's tempting to treat DSR as purely a multiple-testing correction, but the second half of the formula is doing real work. Two strategies with identical ŜR and identical N can have very different DSR if their return distributions differ. A strategy with strongly negative skew (frequent small wins, occasional large losses — the typical shape of short-premium options strategies) requires a larger gap between ŜR and SR₀ to reach the same confidence level as a strategy with symmetric returns.

A worked example from the literature: a strategy with an annualized Sharpe ratio of 2.5, selected after 100 trials with a Sharpe ratio variance of 0.5 across those trials, skew of −3, and kurtosis of 10, over a 1,250-observation backtest, produces a DSR corresponding to roughly a 90% probability that the true Sharpe ratio is positive — which also means a 10% chance the strategy doesn't make money at all, despite an apparently strong headline number.

What DSR requires you to track

The formula only works if you've recorded the inputs honestly:

  1. N — every trial, not just the winner. This includes parameter sweeps, manual iterations, and any variant you tested and discarded, whether or not it was a "formal" backtest run.
  2. σ — the dispersion of Sharpe ratios across those trials, not just the value of the best one.
  3. ŜR, T, skew, and kurtosis of the actual returns of the strategy you ultimately selected.

In practice, most retail backtesting workflows fail at step 1 — traders iterate on a strategy interactively, adjusting a stop-loss or a strike-selection rule by feel, without logging the discarded versions as trials. DSR can't be computed retroactively without that log, and a Sharpe ratio reported without it should be read with the understanding that the true N is almost certainly larger than the one variant being shown.

What DSR doesn't tell you

DSR corrects for selection bias and non-normality at a single point in time. It says nothing about whether a strategy's performance is stable across different market regimes — that's a separate question, addressed by out-of-sample validation methods like walk-forward testing or combinatorial purged cross-validation, not by DSR itself. DSR also assumes the N trials are independent; strategies that are minor variations of each other (e.g., the same signal with strike offsets of 5Δ vs. 10Δ vs. 15Δ) are correlated, not independent, which means the naive formula understates how aggressive the search really was. In that case, the effective number of independent trials is smaller than the literal count, and DSR computed on the literal N will be too conservative — though in the opposite direction of the failure mode it's designed to catch, so it errs safe rather than dangerous.

Quantropy AI is a backtesting platform for retail options traders — describe a strategy in plain English, run it against tick-level historical data, and deploy it, no coding required. Join the waitlist at quantropy.ai.

Sources

  1. Bailey, D. H., & López de Prado, M. (2014). "The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality." The Journal of Portfolio Management, 40(5), 94–107. https://www.davidhbailey.com/dhbpapers/deflated-sharpe.pdf
  2. "Deflated Sharpe Ratio (DSR)." Folio Lab. https://www.foliolab.ai/docs/metrics/deflated-sharpe-ratio
  3. "How to detect false strategies? The Deflated Sharpe Ratio." https://marti.ai/qfin/2018/05/30/deflated-sharpe-ratio.html
  4. "Sharper Sharpe Ratio?" CXO Advisory. https://www.cxoadvisory.com/big-ideas/sharper-sharpe-ratio/
  5. Bailey, D. H., & López de Prado, M. (2012). "The Sharpe Ratio Efficient Frontier." Journal of Risk, 15(2), 3–44.

Read more