TradehouseBuyer guides & toolsPricingPart of TradeHouseTradeHouse Fidelity Lab · Buyer resources
Deflated Sharpe ratio calculator
Enter the figures for the strategy you are considering and the calculator returns the probability that its true Sharpe ratio exceeds what you would expect from simply picking the best of N tries. Everything runs in your browser; nothing you type is sent anywhere.
What each input means
- Observed Sharpe per observation. Mean excess return divided by standard deviation, at the same frequency as your data. For a daily series with an annualised Sharpe of 1.59, enter about 0.10. Do not enter the annualised figure.
- Number of return observations. The count of periods in the sample. More observations tighten the estimate.
- Effective independent trials. How many genuinely different strategies you tried. Near-duplicates count as fractions. If you ran one pre-specified rule, enter 1.
- Standard deviation of Sharpe across trials. How widely the Sharpe ratios of your variants spread, at the same frequency.
- Skewness and kurtosis. Of the strategy returns. Normal kurtosis is 3, not 0. Negative skew and fat tails widen the uncertainty.
Reading the output
The result has two parts. The selection benchmark is the Sharpe you would expect from the best of N unskilled trials. The probability is the chance, under the Bailey and López de Prado approximation, that your observed Sharpe exceeds that benchmark after accounting for sample length, skew and kurtosis. Roughly, 95% or higher is the range many practitioners want before taking a result seriously; below 50% the result is no better than selection luck.
Worked examples
| Observed Sharpe | Observations | Trials | Dispersion | Skew, kurtosis | Result |
|---|---|---|---|---|---|
| 0.10 | 252 | 1 | 0.05 | 0, 3 | 94.3% |
| 0.10 | 252 | 10 | 0.05 | 0, 3 | 63.2% |
| 0.10 | 252 | 100 | 0.05 | 0, 3 | 33.8% |
| 0.10 | 252 | 1,000 | 0.05 | 0, 3 | 16.1% |
| 0.15 | 252 | 100 | 0.05 | 0, 3 | 64.4% |
| 0.10 | 252 | 100 | 0.05 | -0.5, 6 | 34.2% |
| 0.08 | 1,260 | 50 | 0.03 | 0, 3 | 66.1% |
The first four rows hold everything fixed except the trial count. The fifth shows that a higher observed Sharpe at the same trial count lifts the probability sharply. The sixth shows that negative skew and fat tails shift the result only slightly here but widen the error. The last row shows five years of daily data with a modest Sharpe; the long sample helps but 50 trials still leave real doubt.
Limits you should know
- The normal CDF is numerically approximated, so results can differ in the last decimal from other implementations.
- Correlated variants require an effective independent-trial count. Counting them all inflates the benchmark; counting one understates it.
- The probability says nothing about future returns, transaction costs or regime change.
- Sharpe and dispersion must be at the same frequency. Do not mix annualised and daily figures.
Converting an annualised Sharpe
If your platform reports an annualised Sharpe, divide by the square root of the number of periods per year to get the per-observation figure: 252 for daily equities, 52 for weekly, 12 for monthly. An annualised Sharpe of 1.0 on daily data is about 0.063 per day. Apply the same conversion to the dispersion across trials, or compute that dispersion directly from the per-observation Sharpe of each variant.
Beyond one number
For a full verdict on a trade file, Fidelity Lab adds cost netting from a per-venue fee schedule, 11 contamination detectors, a frozen-rule walk-forward and Benjamini-Hochberg control. The first report each month is free. Method: Bailey and López de Prado, The Deflated Sharpe Ratio (2014), equations 1 to 3. A side-by-side with the unadjusted ratio is on deflated vs plain Sharpe.
Frequently asked questions
What number of trials should I enter?
Your honest count of independent strategies tried. If in doubt, run it at your count and at ten times your count.
Why does more data sometimes lower my result?
When the observed Sharpe is below the selection benchmark, a longer sample makes that shortfall more certain.