Crypto deflated Sharpe ratio: test the selected winner

The crypto deflated Sharpe ratio assesses a selected historical Sharpe ratio against a benchmark raised for multiple trials, while allowing for sample length, skewness and kurtosis. It asks whether the winner clears a selection-adjusted statistical hurdle under stated assumptions. The result is a probability-style statistic, not a reduced annual return or a probability of making money in live trading.

A convincing calculation needs information about the search that produced the winner. If the discarded candidates are missing, the displayed Sharpe ratio cannot supply their number or their dependence. Start with the existing backtest trial ledger; this article develops the adjustment that the ledger can help support.

Establish the return series before deflating it

Use a declared series of equal-length, net portfolio returns and a same-period excess-return benchmark. Record the valuation currency, sample dates, missing-period treatment, costs and number of observations. Daily account returns and individual trade returns answer different questions, especially when positions overlap or capital sits idle.

For a multi-asset strategy, reconcile portfolio return contributions before computing the statistic. A return history assembled from the best-performing asset each day can introduce selection inside the portfolio series itself. A correction applied to the final ratio does not make that construction investable.

The numerator uses an arithmetic mean of periodic excess returns. Replacing it with CAGR changes the input statistic; the arithmetic versus geometric mean guide explains that distinction. The Sharpe and Sortino comparison covers the basic denominator conventions.

Keep the selected Sharpe and the other candidates' Sharpes in the same, unannualized sampling unit. William Sharpe's original discussion explains that time scaling depends on the return aggregation assumptions and becomes more complicated with compounding or serial correlation. Attaching a daily observation count to an annualized Sharpe inside the formula creates a different calculation.

The benchmark represents a lucky winner

Bailey and López de Prado's original Deflated Sharpe Ratio paper combines a multiple-trial benchmark with the probabilistic Sharpe calculation. For its zero-mean null, build the approximate benchmark in two steps.

Calculate two standard-normal quantiles: q1 = Φ⁻¹(1 − 1/N) and q2 = Φ⁻¹(1 − 1/(N × e)). The approximate hurdle is B = v × ((1 − g) × q1 + g × q2).

Here N is the number of independent trials, v is the standard deviation of the trials' estimated Sharpes, g ≈ 0.5772157, e is Euler's number and Φ⁻¹ is the inverse standard normal cumulative distribution function. This large-trial-count approximation models the maximum of independent normal trial estimates; small trial counts can add approximation error. It does not require the selected strategy's returns to be normal.

The distinction between two standard deviations matters. The selected strategy's return deviation enters its ordinary Sharpe. The spread of candidate Sharpe estimates enters v. Supplying the first in place of the second gives the benchmark the wrong input.

The displayed approximation requires N > 1. For one genuinely unselected trial, the expected single estimate under this zero-mean null is zero; handle that case separately. Do not send N = 1 through an inverse-normal expression containing zero. Nor should a researcher label an undocumented search a single trial just because only one configuration survived.

Account for uncertainty in the selected Sharpe

Using Equation 2 of the DSR paper, the second stage compares the selected per-period Sharpe S with B:

First calculate Q = 1 − skew × S + ((kurtosis − 1)/4) × S². Then form the normal score z = (S − B) × sqrt(T − 1) / sqrt(Q) and calculate DSR = Φ(z), where Φ is the standard normal cumulative distribution function.

T is the selected series' observation count and must exceed one; a one-observation sample cannot supply an ordinary sample Sharpe. skew and kurtosis describe that same series; kurtosis here is the ordinary standardized fourth moment, equal to three for normal returns. It is not excess kurtosis. Record the moment estimators used, including any small-sample corrections, so another implementation can reproduce the inputs.

The underlying probabilistic Sharpe research considers uncertainty from track-record length and non-normality, including negative skewness and positive excess kurtosis. Its stated sampling condition also matters: increasing frequency is useful only while the independent and identically distributed assumption remains defensible.

Two separate dependence questions therefore remain. Candidate strategies can share the same market moves, affecting the independent-trial count. Returns within one candidate can depend on earlier returns, affecting the validity of the simple uncertainty formula. Correcting one does not correct the other.

The expression is an approximation. Small samples, unstable higher moments or a changing return distribution can undermine its interpretation even when the arithmetic is correct.

A hypothetical calculation changes the conclusion

Consider an illustrative research report with 730 daily observations, selected daily Sharpe S = 0.15, skewness −1.5, ordinary kurtosis 8, and cross-trial Sharpe standard deviation v = 0.04. These are invented summary inputs, not estimates from a crypto dataset. Assume the formula's sampling conditions hold solely to demonstrate its arithmetic.

With N = 100 independent trials, the bracketed maximum factor is approximately 2.530603. The benchmark is therefore 0.04 × 2.530603 = 0.101224 per day.

For these inputs, Q = 1 − (−1.5 × 0.15) + ((8 − 1)/4) × 0.15² = 1.264375. The uncertainty denominator is therefore sqrt(1.264375).

The normal score is (0.15 − 0.101224) × sqrt(729) / sqrt(1.264375), approximately 1.171200. Applying the standard normal cumulative distribution gives DSR ≈ 0.879241, or 87.92%.

Now hold every input except the independent-trial count fixed:

A hypothetical review policy requiring 95% would reach different decisions at 25 and 100 trials. That cutoff is an illustrative research rule, not a capital-allocation recommendation. The table makes uncertainty about N visible instead of disguising it in one precise percentage.

Against a zero benchmark, the same selected series gives a probabilistic Sharpe score of approximately 99.9842%. The gap from 87.92% comes from changing the benchmark to reflect selection; no trade or return was removed.

For another controlled comparison, set skewness to zero and kurtosis to three while retaining N = 100 and all other inputs. The score becomes approximately 90.48%. For the positive S − B in this example, the original negative skewness and larger kurtosis increase the denominator and reduce the score. That directional statement should not be generalized to every possible numerator sign.

Count independent trials without inventing precision

The DSR paper's appendix distinguishes actual runs from implied independent trials and discusses average-correlation approaches, with limitations. Its warning matters operationally: correlated variants do not automatically provide as many independent chances as the raw run count suggests.

Suppose a hypothetical ledger contains 200 parameter combinations. Almost identical lookbacks may produce nearly identical positions. Other variants may trade different instruments but still share a common market exposure. Neither the parameter distance nor the number of ticker symbols directly determines an independent-trial count.

Preserve aligned candidate return series, candidate identifiers and the full selection history. Estimate dependence with a documented method, retain the resulting uncertainty and show sensitivity across justified alternatives. A single average correlation can conceal clusters: one strongly redundant family and another distinct family can average into a misleadingly simple description.

Do not substitute a portfolio's effective asset count or a return series' effective observation count for the independent-trial count. These quantities describe different objects. Similarly, fewer effective trials do not authorize increasing T or deleting unfavorable observations.

When earlier experiments are unavailable, write the limitation directly: “The documented grid contains 200 variants; earlier manual searches are unknown.” A sensitivity table can explore that gap, but cannot recover the missing history. Choosing the smallest plausible count because it passes a threshold creates another selection decision.

Turn the result into an auditable research decision

Save a compact calculation receipt alongside the report:

Reject a calculation as incomplete when its units disagree, required inputs are missing, or the variance expression is nonpositive. Investigate the cause instead of clipping a broken denominator into a plausible score. Check whether one extreme observation dominates the moment estimates; removing it after seeing the answer is a new analysis that needs its own record.

Keep model defects separate from statistical uncertainty. Missing fees, stale marks, impossible fills and information unavailable at the decision time require repairing the return series. A high DSR cannot validate those assumptions, establish market capacity or show that future conditions will resemble the fitted history.

Freeze the research rule before the next evaluation. Preserve the receipt and test the complete selection procedure on later, previously unused information. The useful next action is to request the candidates and their common return clock: without those, the most impressive DSR decimal remains difficult to audit.