What is the deflated Sharpe ratio?
The deflated Sharpe ratio is the probability that a backtest's observed Sharpe ratio exceeds what the best of all the variants tried would have produced by chance, given how many were tried, how long the record is and how non-normal the returns are.
The conventional Sharpe ratio answers a narrower question: how large is the average return relative to its volatility. It is a fine description of a single record. It becomes misleading when the record was chosen from among many, because the chosen one is the one on which chance was most generous, and the ratio includes that generosity.
The deflation works by estimating what the highest Sharpe ratio among a given number of skill-free trials would be, and then asking whether the observed ratio is meaningfully above that benchmark. A strategy that clears the bar has a result that is unlikely to be explained by selection alone. A strategy that does not is indistinguishable from the best of a set of random tries, however good its number looks on its own.
The output is a probability rather than a ratio, which is the right shape for the question. The question was never how high the number is; it was whether the number means anything.
Why does the number of trials matter so much?
The number of trials matters because the expected maximum of many chance results grows with the count, so the more variants that were tried, the higher a result can be while still being entirely explained by luck.
The growth is slower than linear but it does not stop. The best of ten random strategies over a few years of data will show a Sharpe ratio that looks respectable. The best of a hundred looks good. The best of a thousand, which is what a modest parameter grid produces without anyone noticing, looks like something worth trading. None of them has an edge, and each looks better than the last purely because more were tried.
This is the mechanism by which honest people publish backtests that do not work. Nobody set out to cherry-pick. They tried a handful of ideas, refined the promising ones, tested a range of settings, and kept the best. Every step was reasonable and the cumulative count of trials ran into the hundreds, at which point the survivor's Sharpe ratio was mostly selection.
The correction has to be applied to the count that was actually incurred, which is why keeping the count is the part of the discipline that matters most.
What inputs does the calculation need?
The calculation needs the number of effectively independent trials, the spread of Sharpe ratios across those trials, the length of the record in observations, and the skew and kurtosis of the returns.
The trial count and the spread across trials together set the benchmark, the expected maximum that chance alone would reach. A larger count raises it; a wider spread of results across the trials raises it further, because the trials that differed most from each other had the most room to get lucky.
The record length sets how precisely the observed ratio is known. A ratio measured over a few dozen trades has a wide uncertainty band, and a wide band means the observed value has to be much further above the benchmark before it counts. The same ratio over several hundred trades is known more tightly and clears the same benchmark more easily.
Skew and kurtosis correct for the fact that trading returns are not normally distributed. A strategy with occasional large losses has fat tails, and fat tails make an observed Sharpe ratio less reliable than the normal-distribution arithmetic assumes. The adjustment widens the uncertainty accordingly, and for strategies with a few very large outcomes the widening is substantial.
How do you count trials honestly?
Count every variant that was run and whose result influenced what was kept, treat correlated variants as fewer than their raw number, and never reset the count when a new idea replaces an old one.
The first rule is the one most often broken, because most trials do not feel like trials. Changing a lookback from twenty to twenty-five and rerunning is a trial. Adding a filter and checking whether it helps is a trial. Testing the same rule on a second instrument to see if it works there is a trial. Each one gave chance another opportunity, and each belongs in the count.
The second rule prevents over-correction. A hundred settings of one parameter, all producing similar results, are not a hundred independent chances; they are a handful. The effective number is closer to the number of distinct result clusters than to the number of runs, and the spread of results across trials is what the calculation uses to estimate it. A grid of near-identical results deflates less than a grid of wildly different ones.
The third rule addresses the reset. The count belongs to the research effort, not to the current idea. A trader who has tried and abandoned twelve strategies this year and is now testing the thirteenth has a count of thirteen plus every variant of each, and the thirteenth's result has to clear the benchmark that count implies.
How does QuantParadox account for selection?
QuantParadox counts the variants tried on a strategy within a research session and compares the best result against what the best of that many chance results would look like, which is the same correction the deflated Sharpe ratio makes, expressed as a selection-luck check beside the out-of-sample result rather than as a single adjusted ratio.
The count is kept by the platform because it is the input people cannot keep themselves. Every rerun after a change to the rule or its parameters is recorded, and the check is reported against that recorded count, so a result that survived thirty variants is graded as the best of thirty rather than as a single test.
The limitation is the count the platform cannot see. Variants tried elsewhere, in a charting package or a spreadsheet or in the trader's head before the strategy arrived, are not in the ledger, and the benchmark is only as honest as the number it is given. A trader who arrives with a heavily pre-selected idea should treat the reported check as a floor on the correction rather than the whole of it.
The out-of-sample split remains the primary defence, and the selection check is the second layer. A result that clears both is a stronger finding than one that clears either, and a result that clears neither is reported as not established regardless of how its raw Sharpe ratio looks.