The deflated Sharpe ratio (DSR), proposed by Bailey and Lopez de Prado (2014), adjusts a strategy's Sharpe ratio for the selection bias of having tried many strategies, for short track records and for non-normal returns. It tells you how likely it is that the Sharpe ratio is real and not luck.
The problem it solves
If you test many strategies and keep the best, the best Sharpe ratio you see is higher than expected for any single honest test. See data snooping.
An illustrative example
With 5 years of daily data, the standard error of an annual Sharpe ratio for a strategy with no skill is about 1 / √5 = 0.45. Among 100 independent useless strategies, the best is expected to sit about 2.5 standard errors above zero, so
| Quantity | Value |
|---|---|
| Standard error (5 years) | about 0.45 |
| Expected best of 100 useless strategies | about 2.5 x 0.45 = 1.1 Sharpe |
A backtested Sharpe ratio of 1.1 from the best of 100 trials is therefore what luck alone can produce. The numbers are approximate and illustrative.
What the DSR accounts for
- Number of trials. More tests mean a higher hurdle.
- Length of history. Short records have noisy Sharpe estimates.
- Skewness and fat tails. Negative skew and fat tails make a given Sharpe ratio less reliable than a normal distribution implies.
Using it
- Record every variant tried, so the number of trials is honest.
- Treat a DSR probability well below 95% as weak evidence.
- Pair it with out-of-sample testing, walk-forward testing and an economic rationale.
It is one tool against overfitting in backtests.
Frequently asked questions
What is the deflated Sharpe ratio?
It is a Sharpe ratio adjusted for the number of strategies tested, the length of the track record and non-normal returns, to estimate whether it is likely to be genuine.
Why is the best backtest's Sharpe ratio inflated?
Because choosing the best of many tests picks up luck as well as skill.
Who proposed the deflated Sharpe ratio?
David Bailey and Marcos Lopez de Prado, in a 2014 paper.
Does a high deflated Sharpe ratio guarantee future returns?
No. It reduces the chance of being fooled by selection bias but does not account for changing markets or costs.