Overfitting and the deflated Sharpe ratio

The selection problem

Try enough variations of a strategy and one will look excellent purely by chance. The best result from a search is a biased estimate of the truth, and the bias grows with the number of trials.

Suppose every variant you try has a true Sharpe of zero. With $N$ independent trials, the expected maximum Sharpe you will observe is approximately

where $Z^{-1}$ is the inverse normal CDF, $\gamma \approx 0.5772$ is the Euler–Mascheroni constant, and $\sigma_{SR}$ is the standard deviation of the Sharpe estimates across trials. For 100 trials this is roughly $2.5\,\sigma_{SR}$; for 1,000, about $3.3\,\sigma_{SR}$.

The deflated Sharpe ratio

Bailey and López de Prado’s deflated Sharpe ratio (DSR) asks: given how many trials you ran, and given the skewness and kurtosis of the returns, what is the probability that the true Sharpe exceeds zero?

Here $\widehat{SR}$ is the per-period Sharpe, $T$ the number of observations, $\hat\gamma_3$ and $\hat\gamma_4$ the skewness and kurtosis, and $SR_0$ the expected maximum Sharpe under the null from the formula above. A DSR above 0.95 is reasonable evidence of a real edge. Run it on your own numbers with the Sharpe significance tool.

Notice the denominator: negative skew and fat tails increase the uncertainty of the Sharpe estimate, so strategies that sell tail risk need even more evidence.

Symptoms of an overfit strategy

• Many parameters relative to the number of trades.
• Oddly specific values (an RSI period of 13, a stop at 2.37 ATR).
• Performance that collapses when a parameter moves a little — a sharp peak rather than a broad plateau in parameter space.
• Great in-sample results, ordinary or negative out-of-sample.
• A narrative invented after seeing the backtest.

Defences

1. Count your trials honestly, including the ones you abandoned, and deflate.
2. Hold out data you never touch until the very end — and only once.
3. Prefer plateaus to peaks: choose parameters from the middle of a broad region of good performance.
4. Fewer degrees of freedom: simpler rules, fewer filters.
5. Validate properly: walk-forward and purged cross-validation.
6. Demand a reason: an economic or behavioural story for the edge, stated before testing.