Tools / Is my Sharpe real?
Is my Sharpe real?
Test a backtest Sharpe ratio for significance and deflate it for the number of strategies you tried.
SR* adjusted for N trials
- t-statistic
- Std error of Sharpe
- 95% interval
- Best of N by luck alone
- Deflated Sharpe (P real)
- Years needed (95%, 1 trial)
Confidence vs how hard you searched. The probability your Sharpe is real, as the number of variants tried grows.
What the numbers mean
Deflated Sharpe (Bailey & López de Prado) is the probability that the true Sharpe is above zero after accounting for the number of variants you tried, the length of the track record, and the skewness and kurtosis of returns. Above 95% is reasonable evidence; below 50% means the backtest is more likely luck than skill.
Best of N by luck is the annualised Sharpe you would expect from the best of N strategies with no edge at all. If your backtest Sharpe is not comfortably above it, your search found noise. Negative skew and fat tails (high kurtosis) widen the uncertainty, so tail-selling strategies need more evidence.
The theory is in the lesson.
Turn this into a tool you own
A spec is not a trading system. Paste this into your coding agent and AgenKit runs a specialist team — architecture, test-first build, risk gates, review — with paper trading on by default.
/agenkit build a strategy research harness that logs every backtest variant tried, computes probabilistic and deflated Sharpe ratios with skew and kurtosis, and refuses to promote a strategy below a configurable confidence- npx agenkit install engineering-kit in your project
- Paste the command into your agent and approve each phase
- Get tests, backtest, risk controls and a paper-trading switch