Deflated Sharpe Ratio
The question it answers
You ran a backtest and got a Sharpe ratio you like. Two things can make that number lie to you. First, it was estimated from a finite, possibly skewed, possibly fat-tailed sample, so the raw Sharpe overstates how confident you should be that the true Sharpe is above zero. Second, and far more dangerous: if you tried more than one strategy variant before picking this one, the best of a pile of mediocre-to-bad variants can look great by chance alone, even if every variant has zero real skill. The Probabilistic Sharpe Ratio (PSR) fixes the first problem. The Deflated Sharpe Ratio (DSR) fixes the second by asking the harder question: given how many things you tried, what's the probability that this Sharpe reflects real skill rather than the best draw from a large pile of noise?
The math, in plain English
PSR asks: what's the probability that the true Sharpe ratio exceeds some benchmark \(SR^*\), given the observed Sharpe, the sample length, and the shape of the return distribution?
\(\Phi\) is the standard normal CDF, \(T\) is the number of return observations, \(\hat\gamma_3\) is sample skewness, and \(\hat\gamma_4\) is sample kurtosis under the convention that a normal distribution scores \(\hat\gamma_4 = 3\) (not the "excess kurtosis" convention where normal scores 0). Negative skew and fat tails widen the denominator and push PSR down: the same observed Sharpe earns less confidence when the return distribution is nastier than normal.
DSR is PSR evaluated against a benchmark that isn't zero: it's the Sharpe ratio you'd expect the best of N zero-skill trials to show by chance. That expected maximum is:
\(\gamma \approx 0.5772\) is the Euler-Mascheroni constant, \(N\) is the number of trials, and \(\hat\sigma_{SR}\) is the standard deviation of the periodic Sharpe ratios across the trials. This is a first-order approximation from extreme value theory: even under a pure luck null (every trial has zero true skill), the best of many draws is expected to look good, and the spread of Sharpes across your trials tells you how good. DSR then plugs this expected-max Sharpe back into the PSR formula as \(SR^*\):
More trials, or more spread in trial quality, raise the bar the chosen strategy has to clear. With a single trial (\(N=1\)) the formula has no multiple-testing benchmark to apply, so skepsis returns \(E[\max SR] = 0\) and DSR collapses to plain PSR — flagged elsewhere as an optimistic special case (see below).
skepsis's implementation reproduces the DSR paper's own published numerical example to all four printed decimals as a golden test — the formula above is exactly the code, not a paraphrase.
When it lies
skepsis deflates with the raw trial count, not the paper's
effective-trials correction. The DSR paper's full method estimates an
effective number of independent trials by clustering correlated variants
together (e.g. via correlation-based clustering of the trial return series)
so that ten near-duplicate parameter settings count closer to one
independent trial than ten. skepsis does not implement that correction: N
in the formula above is literally the number of columns you pass in
trials=. This is a deliberate design choice (spec §3.2.8), and it is
conservative — clustering correlated trials down to a smaller effective N
would raise \(E[\max SR]\) less, which raises DSR. Skipping the correction
means skepsis's DSR is always at least as harsh as, and often harsher than,
the paper's effective-N version. If your 50 "trials" are actually five
strategies each swept across ten near-identical parameter values, expect
skepsis to deflate more than the textbook method would. That's the
trade-off: no hidden clustering heuristic to get wrong, at the cost of
sometimes being pessimistic about genuinely correlated trial families.
Heavy autocorrelation strains the assumptions. The \(\sqrt{T-1}\) scaling and the skew/kurtosis correction come from an asymptotic, close-to-IID argument. Returns with strong serial correlation (trending strategies on trending assets are the classic case) violate that, and the effective sample size is smaller than \(T\) suggests. skepsis surfaces this indirectly: when the bootstrap's auto-selected block length (Politis-White) comes out large, it's a sign the series is far from IID, and skepsis emits a warning that PSR/DSR should be read with extra skepticism.
A single trial gives you no real deflation. If you don't pass trials=,
skepsis runs DSR with \(N=1\), which is mathematically identical to plain
PSR against a zero benchmark. It says so explicitly, with a warning: this
number is almost certainly optimistic, because you tried more than one
thing to get here even if you didn't record the others.
Worked example
The "Deflating the Golden Cross" demo evaluates a moving-average crossover family on 11 years of BTC-USD, sweeping 34 fast/slow window combinations.
Act 1 — raw returns. The best combination, f5_s100, posts an
annualized Sharpe of 1.464 against a buy-and-hold Sharpe of 1.064
over the same period. Fed the full 34-trial sweep, DSR does not kill this
result — the verdict comes back MODERATE, with the only flagged rule
being PBO (0.448, covered in the PBO explainer). The reason DSR
survives is that the returns are real: every one of the 34 variants clears
a Sharpe of 1.05, because they're all long BTC through a decade-long bull
run. High DSR here does not mean "real timing skill" — it means the
deflation correctly recognizes that this family of strategies would have
looked good almost regardless of which fast/slow pair you'd picked, because
the edge is beta, not selection.
Act 2 — excess over buy-and-hold. A timing strategy earns its keep only if it beats doing nothing, so evaluate each variant's return minus buy-and-hold. Only 1 of the 34 trials shows positive excess Sharpe, and the best of them manages just 0.084. Deflated against 34 trials, that becomes a DSR of 0.345 — worse than a coin flip that the true excess Sharpe is even positive, let alone attractive — and the verdict flips to LIKELY_OVERFIT. This is the honest version of the question: not "does this beat zero," but "does this beat the trivially available alternative, once you account for how many ways you tried to beat it."
References
- Bailey, D. H., & López de Prado, M. (2012). "The Sharpe Ratio Efficient Frontier." Journal of Risk, 15(2).
- Bailey, D. H., & López de Prado, M. (2014). "The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality." Journal of Portfolio Management, 40(5).
- Sharpe, W. F. (1994). "The Sharpe Ratio." Journal of Portfolio Management, 21(1).