Skip to content
QuantCalc Tools
Directory

Cagr Irr Drawdown Compounding

Probabilistic & Deflated Sharpe Ratio (DSR): Backtest Multiple-Testing Math

Testing 250 strategy variations on 5 years of random noise guarantees an expected maximum backtest Sharpe Ratio above 1.45 purely by luck; master exact PSR, DSR, and MinTRL equations.

By Elena Rostova, FRM, CQFReview not recorded · Updated
Probabilistic & Deflated Sharpe Ratio (DSR): Backtest Multiple-Testing Math — QuantCalc Tools 2026 quantitative & supervisory data card
QuantCalc Tools (CAGR vs IRR, Volatility Drag, Sharpe/Sortino & Drawdown Recovery): Bailey & Lopez de Prado multiple-testing Extreme Value hurdle and Mertens tail SE.Source: QuantCalc Tools Institutional Research DeskLicense: CC-BY-4.0

Modern cloud clusters let a quantitative researcher sweep $K = 500$ moving-average windows, RSI thresholds, and ML hyperparameters before lunch, pick the single equity curve with a $2.10$ annualized Sharpe Ratio, and pitch it to an investment committee. Without adjusting for multiple-testing selection bias (data snooping) and non-normal higher moments, that $2.10$ backtest is statistically indistinguishable from pure white noise! David Bailey and Marcos Lopez de Prado (2012, 2014) solved this crisis in empirical finance by introducing the Probabilistic Sharpe Ratio (PSR), Minimum Track Record Length (MinTRL), and Deflated Sharpe Ratio (DSR). Benchmark raw ratios in our Sharpe & Sortino Ratio Calculator, review Lo (2002) Sharpe statistics, and anchor in our CAGR, IRR & Drawdown Hub.

Exact Equations: PSR, MinTRL, Expected Maximum Sharpe $\text{SR}_0$, and DSR#

Work strictly in non-annualized sample frequency ($N$ observations, non-annualized sample Sharpe $\widehat{\text{SR}}$, skewness $\hat{\gamma}3$, and **Pearson kurtosis $\hat{\gamma}4$ where Normal $= 3.0$—never confuse Pearson $\gamma4 = 3$ with excess kurtosis $0$, or the $((\gamma4 - 1)/4)\widehat{\text{SR}}^2$ term goes wrong!). First, PSR** computes the probability that true skill exceeds benchmark hurdle $\text{SR}^$ given non-normal skew/kurtosis. Second, DSR* sets that hurdle $\text{SR}_0$ equal to the Expected Maximum Sharpe Ratio across $K$ independent zero-skill trials via Extreme Value Theory (Euler-Mascheroni constant $\gamma \approx 0.57721566$):

Formula or model from the articleNot independently verified
Bailey & Lopez de Prado (2012, 2014) Exact PSR, MinTRL & Deflated Sharpe Ratio (DSR):
  • Non-Normal Variance Weight : Vmertens = 1 - gamma3 * SRhat + ((gamma4 - 1) / 4) * SRhat^2 [Pearson gamma4 = 3 for Normal!]
  • 1. Probabilistic Sharpe : PSR(SR) = Phi( [ (SR_hat - SR) * sqrt(N - 1) ] / sqrt(V_mertens) )
  • 2. Min Track Record Length : MinTRL(alpha) = 1 + Vmertens * [ z{1-alpha} / (SR_hat - SR*) ]^2 (in sample periods!)
  • 3. Expected Max Sharpe across K Independent Zero-Skill Trials (Variance of Trial SRs = Var[{SR_k}]):
  • SR0 = sqrt( Var[{SRk}] ) * [ (1 - gamma) * Phi^(-1)(1 - 1/K) + gamma * Phi^(-1)(1 - 1/(K * e)) ]
  • where Euler-Mascheroni gamma = 0.57721566, e = 2.7182818, and under IID unit trials Var[{SR_k}] ~= 1/N (or q/N annualized)
  • 4. Deflated Sharpe Ratio : DSR = PSR( SR* = SR_0 ) [Require DSR >= 0.95 to reject data-mining null!]
  • Worked Backtest Overfitting Audit (5 Years Monthly N = 60, Reported Annualized SR = 1.80 -> Monthly SR_hat = 0.5196):
  • Assume K = 200 parameter trials tested (Annualized std across trials = 0.45 -> Monthly std = 0.1299),
  • Skew gamma3 = -1.20, Pearson Kurtosis gamma4 = 6.50:
  • 1. Expected Max Lucky Monthly Hurdle SR_0:
  • Phi^(-1)(1 - 1/200) = 2.5758; Phi^(-1)(1 - 1/543.66) = 2.9045
  • Bracket = 0.42278(2.5758) + 0.57722(2.9045) = 2.7655
  • Monthly Hurdle SR_0 = 0.1299 * 2.7655 = 0.3592 (Equivalent to 1.244 Annualized Sharpe purely from luck!)
  • 2. Non-Normal Denominator V_mertens = 1 - (-1.20)(0.5196) + (5.50/4)(0.2700) = 1 + 0.6235 + 0.3713 = 1.9948
  • 3. DSR = Phi( [ (0.5196 - 0.3592) * sqrt(59) ] / sqrt(1.9948) ) = Phi( 1.2320 / 1.4124 ) = Phi(0.872) = 80.8%!
  • Verdict: FAILS 95% institutional gate (80.8% < 95.0%)! Overfitted across 200 trials with negative skew!

Look at how devastating that audit is: an annualized Sharpe of $1.80$ over 5 years has a naive Gaussian single-trial $z$-score of $1.80 \times \sqrt{5} = 4.02$ ($p = 0.00003$, 99.997% confidence!), yet once we account for sweeping $K = 200$ parameter combinations and negative skew $\gamma_3 = -1.20$, confidence plunges to 80.8% (p = 0.192)—failing institutional deployment gates!

Four-stage Deflated Sharpe Ratio backtest validation pipeline counting effective independent trials K_eff and computing DSR confidence
Four-stage Deflated Sharpe Ratio backtest validation pipeline counting effective independent trials K_eff and computing DSR confidenceVerified Exhibit

Expected Maximum Sharpe Ratio of Pure Noise Across $K$ Trials ($T = 5$ Years IID)#

Under $T = 5$ years of zero-skill Gaussian white noise (annualized trial standard deviation $\sigma{\{\text{SR}\}} = 1/\sqrt{5} = 0.4472$), watch how rapidly the **Expected Maximum Backtest Sharpe Ratio $\mathbb{E}[\maxk \widehat{\text{SR}}_k]$** climbs purely by luck as a researcher tests more variations:

Extreme-Value Noise Floor: Expected Maximum Sharpe Across K Backtest Trials#

Quantify the exact annualized Sharpe Ratio generated purely by Gaussian luck across parameter grid sweeps:

Quantitative Distribution · Extreme-Value Multiplier ($Z_{\max}$)
Extreme-Value Multiplier ($Z_{\max}$)Expected Max Noise Sharpe ($T = 5\text{y}$)
01$K = 1$ (True Pre-Registere…$0.00\,\sigma$02$K = 10$ (Small Parameter G…$1.54\,\sigma$03$K = 100$ (Feature Selectio…$2.51\,\sigma$04$K = 1{,}000$ (ML / AutoML …$3.24\,\sigma$05$K = 10{,}000$ (Genetic / D…$3.85\,\sigma$
Independent Strategy Trials Tested ($K$)Extreme-Value Multiplier ($Z_{\max}$)Expected Max Noise Sharpe ($T = 5\text{y}$)Expected Max Noise Sharpe ($T = 2\text{y}$)Minimum 5y Backtest SR Needed for DSR = 95% (Normal)
$K = 1$ (True Pre-Registered Hypothesis)$0.00\,\sigma$$0.00$$0.00$$0.74$
$K = 10$ (Small Parameter Grid)$1.54\,\sigma$$0.69$$1.09$$1.45$
$K = 100$ (Feature Selection Sweep)$2.51\,\sigma$$1.12$$1.77$$1.91$
$K = 1{,}000$ (ML / AutoML Grid Search)$3.24\,\sigma$$1.45$$2.29$$2.27$
$K = 10{,}000$ (Genetic / Deep Alpha Mining)$3.85\,\sigma$$1.72$$2.72$$2.58$
Expected maximum annualized Sharpe ratio generated purely by random white noise over a 5-year backtest across 10 to 10000 trials
Expected maximum annualized Sharpe ratio generated purely by random white noise over a 5-year backtest across 10 to 10000 trialsVerified Exhibit

Estimating Effective Independent Trials ($K_{\text{eff}}$) When Sweep Grids Are Correlated#

What if those $200$ backtested variations were moving averages from MA(10) to MA(209) whose return streams are $92\%$ correlated with one another? Treating $K = 200$ as $200$ independent trials would over-penalize the strategy! Lopez de Prado & Lewis (2019) resolve correlated grids by clustering the $K \times K$ trial correlation matrix (or using average pairwise correlation $\bar{\rho}$ via $K{\text{eff}} \approx 1 + (K - 1)(1 - \bar{\rho})$): if $200$ sweeps have $\bar{\rho} = 0.85$, effective independent trials equal $K{\text{eff}} = 1 + 199(0.15) \approx 31$, giving an exact, fair selection-bias hurdle!

Combining DSR with Lo Autocorrelation and Out-of-Sample Paper Verification#

Before allocating live margin on FIX API or MT5 infrastructure, chain both Lo (2002) serial-correlation unsmoothing, Information Ratio breadth checks, AND Bailey-Lopez de Prado DSR, verify skewness/kurtosis tail exposure via Cornish-Fisher Modified VaR, and review our transparent testing protocols in methodology and verify.

Frequently asked questions

What is the #1 coding bug engineers make when implementing PSR and DSR in Python?

Passing pandas/scipy default Fisher EXCESS kurtosis (where Normal = 0) into ((gamma_4 - 1) / 4) * SR^2! Bailey & Lopez de Prado wrote (gamma_4 - 1)/4 using PEARSON non-excess kurtosis (where Normal = 3, so (3-1)/4 = 0.5). If using scipy excess kurtosis k_excess, you MUST write ((k_excess + 2) / 4) * SR^2!

Can I plug annualized Sharpe Ratios directly into the PSR and DSR formulas?

No! Mertens’ higher-moment variance formula V = 1 - gamma_3*SR + ((gamma_4 - 1)/4)*SR^2 is strictly scale-specific to the observation frequency at which N, skewness gamma_3, and kurtosis gamma_4 are measured. Convert annualized SR back to per-period SR_hat = SR_ann / sqrt(q) first!

How do I calculate Minimum Track Record Length (MinTRL) for a live emerging manager?

Given monthly SR_hat = 0.35 (annualized 1.21) against benchmark SR* = 0, skew = -0.8, and Pearson kurtosis = 5.0, V_mertens = 1 + 0.28 + 0.1225 = 1.4025. At 95% confidence (z = 1.6449), MinTRL = 1 + 1.4025 * (1.6449 / 0.35)^2 = 31.99 months (~2.67 years minimum track record).

Does Combinatorial Purged Cross-Validation (CPCV) replace the Deflated Sharpe Ratio?

They are complementary! DSR tests whether the best point-estimate Sharpe beats the multiple-testing noise floor analytically, whereas CPCV generates an empirical distribution of out-of-sample paths to measure the Probability of Backtest Overfitting (PBO) directly.

Publication record

Published
13 November 2026
Updated
14 November 2026
Author
Elena Rostova, FRM, CQF
Review
Review not recorded
Automated fact check
Passed in the stored pipeline record
Source URLs
4 recorded

AI assistance recorded: anthropic/claude-opus-4.6. This field does not establish human review or source verification.

Academic & Institutional Attribution

Cite This Research & Embed Live Data Card

Open Attribution License (CC BY 4.0)

Standard academic & institutional research paper citation format.

Elena Rostova, FRM, CQF. (2026, November 14). Probabilistic & Deflated Sharpe Ratio (DSR): Backtest Multiple-Testing Math. QuantCalc Tools. https://quantcalctools.com/guides/probabilistic-deflated-sharpe-ratio-multiple-testing-backtest-bias
Explore Cagr Irr Drawdown Compounding Hub (20) →

QuantCalc Tools Data, Benchmarks & Contributor Records

Probabilistic & Deflated Sharpe Ratio (DSR) Formula Guide | QuantCalc Tools