Deflated Sharpe Ratio

Deflated Sharpe Ratio gives the deflated Sharpe ratio its first chart-native implementation. Per-bar returns are reduced to their first four moments over an expanding or rolling sample; the pane plots the probability that the observed Sharpe ratio beats what the best of N unskilled trials would reach by luck. A Probabilistic Sharpe Ratio line tests a fixed benchmark instead; the Selection Penalty fill between the two is the cost of picking your best backtest.

How to Trade the Deflated Sharpe Ratio?

• DSR at or above the significance threshold: the Sharpe estimate survives the best-of-N luck hurdle.
• DSR near 0.5 or lower: the result sits where the luckiest of many random configurations would land.
• High PSR, low DSR: returns look genuine against a fixed benchmark yet fail the luck hurdle — the trial count is doing the damage.

Point the Source at an equity curve or another indicator's output, and read it alongside the Library's validation toolkit. Alerts fire on threshold crossings and a sink below 0.5.

Deflated Sharpe Ratio Settings

• Source (default close) and Returns (default Percent): tested series and return type (Log, or Difference for P&L series).
• Sample Window (default Expanding) with Rolling Length (default 252): whole history or a moving sample.
• Minimum Observations (default 30): required before outputs show.
• Risk-Free Rate (Annual %) (default 0) and Periods Per Year (0 = Auto) (default 0): annualization details.
• Number of Trials (N) (default 10): effectively independent configurations examined.
• Trial SR Dispersion (Annualized) (default 0.5): Sharpe spread across trials.
• Significance Threshold (default 0.95): the significance cut.
• Show Probabilistic Sharpe Ratio (default on) with PSR Benchmark SR (Annualized) (default 0).
• Show Dashboard (default on) plus location, size, colors.

Frequently Asked Questions

What does this add over the plain Sharpe Ratio?

The Sharpe Ratio is a point estimate that says nothing about how it was obtained. The DSR turns it into a probability accounting for estimation noise on a finite, non-normal sample and best-of-N selection.

How should I choose N?

Count every configuration you seriously evaluated — sweeps, variants, abandoned ideas. Understating the count is the classic way this test is defeated; correlated attempts justify a smaller effective N.

Why is the plot empty at the left of the chart?

Skewness and kurtosis estimates are unreliable on tiny samples, so nothing is drawn until Minimum Observations valid returns exist. In rolling mode the requirement is capped at the window length.