Is your Sharpe real,
or the best of many tries?
A Sharpe of 1.4 is a claim. A Sharpe of 1.4 that is the best of 200 attempts is a different claim wearing the same number. Enter three figures and this page reports the probability the true Sharpe is above zero — and how many years of data you actually need.
Test a Sharpe
Count every version you tested and discarded: parameter sets, filters, symbols, date ranges. If you cannot count them, the result cannot be deflated. Nothing is sent anywhere — the arithmetic runs in this tab.
| If you tried this many | P(true Sharpe > 0) |
|---|
Paste this into a forum thread
The probability is the deflated Sharpe ratio: the reported Sharpe minus the Sharpe you would expect from the best of that many random attempts, converted to a normal probability. The years figure is the shortest record at which that probability would reach 95% — the same test, run forward.
Enter a Sharpe, a record length and a trial count, then press Calculate. Press Load example to see a typical vendor claim fall apart.
What this answers
One question: given how hard you looked, is this Sharpe distinguishable from luck? It is not a forecast, a rating, or a buy signal. It is the arithmetic that most backtest write-ups skip, because the input that decides the answer is the one nobody wants to publish — how many things were tried before this one.
The example, worked through
Load the example: a Sharpe of 1.40 over 2 years, chosen as the best of 200 variations. That is a strong, plausible-sounding claim. The tool returns about 22% — roughly a one-in-five chance the true Sharpe is above zero. The other four times out of five, a strategy with no edge would have produced something at least this good, given 200 attempts. It also says you would need about 9.9 years of data before that same Sharpe reached 95% confidence. The backtest has two.
The same 1.40 over the same 2 years, tried once, is a different story entirely. Only the trial count changed. That is the whole point of the page: the number on the marketing page is not the number that matters.
Why the trial count decides everything
Test enough zero-edge strategies and the best one will look excellent. The expected best Sharpe grows with the number of attempts, and it grows faster than intuition suggests — roughly like the square root of twice the log of the trial count. At 200 trials on two years of data, the benchmark Sharpe a random search would reach is already around 1.5. A reported 1.4 is below what noise typically delivers at that search intensity.
This is why “our AI found the optimal settings” is an evasive answer rather than a strong one. The optimal settings are what an optimiser is for, and the optimiser will always find them, edge or no edge.
How the arithmetic works
The probability is the deflated Sharpe ratio (Bailey and López de Prado, 2014). It takes the reported Sharpe, subtracts the expected maximum Sharpe from that many independent trials, and scales the difference by the standard error of a Sharpe estimate over the record length. The years figure is the minimum track record length at 95% confidence, solved numerically: the benchmark Sharpe shrinks as the record grows, so the two figures are computed together and always agree.
The closed form here gives 0.384 for a Sharpe of 1.0 over five years as the best of 100 trials. The essay’s Monte Carlo run of the same setup gave 0.383. They are the same answer at the precision that matters.
What this does not tell you
A high probability is not a pass. This calculation knows nothing about transaction costs, look-ahead bias, survivorship bias, regime change, or capacity. A strategy can be statistically distinguishable from selection noise and still lose money on every trade after spread.
It also assumes the trials were roughly independent and that the returns are close to normal. Real strategies have fatter tails, which makes the honest probability lower than the number shown, not higher. Treat this as a ceiling on confidence, not a floor.
For the costs and leakage checks, run the Lie Detector on the prediction file. For sample size on a win rate, use the Wilson calculator. For the selection problem across many strategies, the CSCV and PBO method is the right test.
Questions
What is a deflated Sharpe ratio?
The deflated Sharpe ratio converts a reported Sharpe into the probability that the true Sharpe is above zero, given how many strategy variations were tried to find it. A Sharpe of 1.0 measured once and a Sharpe of 1.0 that is the best of 100 attempts are the same number and completely different claims.
How many years of track record do I need for a Sharpe of 1?
About 2.7 years of daily data at 95% confidence, if the strategy was the only one tried. As the best of 100 variations, the same Sharpe needs about 17 years. As the best of 1,000, about 24. The reported number does not change; the evidence behind it does.
Why does the number of trials matter so much?
Because the best of many random strategies will always look good. Testing 100 zero-edge strategies and keeping the winner produces an impressive Sharpe from pure noise. The deflated Sharpe ratio subtracts the Sharpe you would expect from the best of that many attempts, which is why the trial count is the input that decides the answer.
What if I do not know how many strategies I tried?
Then the result cannot be deflated, and any p-value attached to it is optimistic. Count every parameter set, every filter, every market you tested on, and every version you deleted. The number is usually far larger than people remember, which is exactly why the answer changes.
Does a high deflated Sharpe prove the strategy is profitable?
No. It says the measured result is unlikely to be selection noise. It does not account for transaction costs, look-ahead bias, survivorship bias, regime change, or capacity. Those are separate checks, and a strategy can pass this one and still lose money.
Is this investment advice?
No. It is arithmetic on a claimed statistic. It does not predict returns, and it says nothing about the next market regime.