TradehouseBuyer guides & toolsPricingPart of TradeHouseTradeHouse Fidelity Lab · Buyer resources
Deflated Sharpe vs plain Sharpe
A plain Sharpe ratio answers one question: given this return series, how much excess return came per unit of volatility? A deflated Sharpe ratio asks a different one: given that you looked at many strategies and kept the best, how likely is it that the best one has a true Sharpe above zero? The plain figure is silent about the search that produced it. The deflated figure is built around it.
The same strategy, scored both ways
Take a daily strategy with an observed Sharpe of 0.10 per day (about 1.59 annualised), 252 observations, zero skew and normal kurtosis. The spread of Sharpe values across all the variants you tried is 0.05. The calculator on the Deflated Sharpe ratio calculator page uses the Bailey and López de Prado formulation, and gives these probability estimates that the true Sharpe is above the selection benchmark:
| Effective independent trials | Selection benchmark, per day | Same, annualised | Deflated Sharpe probability |
|---|---|---|---|
| 1 | 0 | 0 | 94.3% |
| 10 | 0.0787 | 1.25 | 63.2% |
| 100 | 0.1265 | 2.01 | 33.8% |
| 1,000 | 0.1628 | 2.58 | 16.1% |
The plain Sharpe is 0.10 in every row. Only the number of trials changed, and the probability that the result is real moves from comfortable to doubtful. At 100 trials the best of the batch is expected to reach a Sharpe of about 0.127 per day by luck, which is already above the 0.10 you observed.
What each ratio is good for
| Plain Sharpe | Deflated Sharpe | |
|---|---|---|
| Inputs | Mean and standard deviation of returns | Sharpe, number of observations, number of independent trials, dispersion of Sharpe across trials, skewness and kurtosis |
| Corrects for selection among many tries | No | Yes, through the expected maximum of N trials |
| Corrects for short samples | No | Yes, standard error grows as observations fall |
| Corrects for fat tails and skew | No | Yes, through skewness and kurtosis in the variance term |
| Output | A ratio | A probability |
| Biggest risk when misused | Overstates a cherry-picked result | Understates if you claim far more trials than you really ran, or forget that correlated variants are not independent |
Counting trials honestly
The input people get wrong is the number of effective independent trials. Fifty parameter values that differ by one decimal place are not fifty independent bets; they are closer to a handful. An honest estimate comes from clustering the variants by how correlated their return streams are. A conservative habit is to run the calculator at your count and at ten times your count. If the verdict only survives the low number, treat it as unproven.
Frequency matters
The Sharpe you enter and the dispersion across trials must use the same observation frequency. Mixing an annualised Sharpe with daily dispersion makes the benchmark meaningless. Convert the annualised figure to per-observation by dividing by the square root of the periods per year before using the calculator.
Where Fidelity Lab fits
The calculator handles one Sharpe at a time and relies on you to supply the trial count. Fidelity Lab's Starter and Pro plans include a Leaderboard Auditor that takes an entire strategy or agent leaderboard, applies the deflated-Sharpe and max-of-N correction across it, and reports a verdict, so the trial count comes from the submission rather than from memory. The free tier and the single $99 audit cover one book at a time.
Neither number predicts future returns. A deflated Sharpe of 90% says the result is unlikely to be pure selection noise given the stated inputs; it does not say the edge will persist, survive a regime change or clear execution costs. That last question is why the full report also nets fees from a per-venue schedule.
Frequently asked questions
Is a deflated Sharpe always lower than a plain Sharpe?
The deflated figure is a probability, not a ratio, so the two are not on the same scale. With a single trial and a decent sample it can be high; with many trials it falls.
What is a good deflated Sharpe probability?
There is no universal cut-off. Many practitioners look for 95% or higher before sizing up, but that threshold is a convention, not a theorem.
Does deflating fix overfitting?
It quantifies one source of it, selection among many trials. Lookahead, bad fees and duplicated rows still need separate checks.