Deflated Sharpe Ratio (DSR)
Discounts the luck of having tried thousands of combinations. It answers: «if I searched so much, is this Sharpe real or was it inevitable that something looked great by chance?».
Imagine you test 10,000 parameter combinations looking for a good algorithm. Even if none had a real edge, by pure statistics one would come out with a spectacular Sharpe — just as if 10,000 people flip coins, one will get 10 heads in a row without their coin being loaded. The Deflated Sharpe Ratio (DSR) corrects exactly that: it penalizes your Sharpe according to how many combinations you tried, to tell a genuine edge from the lottery prize of mass sampling.
The DSR comes from the research of Marcos López de Prado, a reference in institutional quantitative trading. TradingNote computes it automatically on every algorithm, so you inherit a fund's rigor without doing the math.
Why the normal Sharpe deceives you when exploring
The Sharpe measures return per unit of risk, and it's a great metric… for ONE system. The problem appears when you use it to pick the best of MANY. The more combinations you try, the higher the «winner's» Sharpe will be even if they're all noise. This is called selection under multiple testing, and it's the mechanism by which most «optimized» algorithms fail in real life.
What goes into the calculation
| Factor | Effect on DSR |
|---|---|
| Number of combinations tried | The more, the more penalty: it was likelier to find a winner by chance. |
| History length (number of trades) | More data = more confidence. A Sharpe over 1,000 trades is worth more than over 50. |
| Skew and kurtosis of returns | Returns with fat tails (big shocks) get more penalty. |
How to read it
The DSR reads as a probability: how likely it is that the observed Sharpe is real and not a product of search luck. The higher, the better. An algorithm with a sky-high backtest Sharpe but a low DSR is a clear red flag: its shine is because you tried too much, not because it has an edge.
Common trap: «my algorithm has Sharpe 3.5, it's amazing». If that 3.5 came from exploring 20,000 combinations, the DSR may reveal it was almost inevitable to find that number by chance. The Sharpe without the context of how many tests is meaningless.
The DSR is the first of the four filters and the cheapest to run. Use it as a first cut: if an algorithm doesn't pass the DSR, don't waste time on Walk-Forward or Multiverse. Whatever survives the DSR, run it through the other three.