The Sharpe ratio compresses a return distribution into two numbers: a mean and a standard deviation. Those two describe the distribution completely if it is normal, and only then.
Financial return distributions are not normal, and the two ways they depart from normal are the two that decide what happens to your money.
Skew says which side the surprises arrive from. Negative skew is the shape of many small gains and rare large losses.
Kurtosis says how heavy the tails are: how often the distribution produces moves that a normal distribution treats as essentially impossible.
The second one can be measured rather than left as an adjective. Here it is on the asset we trade: 3,241 daily returns of ether, covering 8.9 years to August 2026, taken from our own fifteen-minute candle history.
Skew is −0.89. Excess kurtosis is 11.9. Bitcoin over the same window gives −0.96 and 15.7, so this is not a quirk of one series.
What that kurtosis does to the tails, against a normal distribution with the same standard deviation:
| move | observed in the series | normal distribution |
|---|---|---|
| 3 standard deviations | once in 66 days | once in 370 days |
| 4 standard deviations | once in 216 days | once in 15,787 days, about 43 years |
| 5 standard deviations | once in 540 days | once in 1.7 million days |
Crypto trades every day, so these are calendar days rather than exchange days throughout.
A four-sigma day arrives about seventeen times per decade rather than once in forty-three years. That is the whole argument about fat tails, expressed as a frequency you can count.
The observed column is a count we made on our own data, so here is a check on it that does not depend on trusting us. A Student-t distribution has excess kurtosis 6/(ν−4), so the measured 11.9 implies ν = 4.5. That ν was fixed by the kurtosis alone and was told nothing whatever about how often large moves occur. Compute its tail probabilities anyway: once in 80, 242 and 600 days for the three thresholds, against the 66, 216 and 540 we counted.
The fit was not asked for and it is not exact. What it means is that a single number, the kurtosis, carries most of the information about how often the extremes arrive, which is precisely the number the Sharpe ratio discards.
Consider what a strategy with negative skew and heavy tails does to the two inputs of the ratio.
The mean is earned steadily, in the ordinary weeks. The standard deviation is also measured mostly in the ordinary weeks, because the extraordinary ones are rare by definition. So the denominator describes the calm part of the distribution while the risk lives in the part it barely samples.
Every quiet year improves the ratio and does nothing to the size of the eventual loss.
That is a mechanical property of the formula, not a claim about anyone's behaviour. Two strategies can report identical Sharpe ratios where one has symmetric noise and the other gives back years of gains in a fortnight, because the ratio has nowhere to put the difference.
The Probabilistic Sharpe Ratio asks a more careful question: given this many observations, with this skew and this kurtosis, what is the probability that the true Sharpe ratio exceeds a threshold you care about?
Two things are folded in. Negative skew and heavy tails widen the uncertainty around the estimate, so an ugly-shaped strategy needs a longer record to support the same claim. And a short record widens it further, for a reason the table above makes concrete.
Take the five-sigma row: once in 540 days on the measured series. A two-year track record covers 730 days, so it should contain one or two such days. Treat the arrivals as random and the chance of a two-year window containing none at all is about one in four.
That window is not a short record of the strategy. It is a record of the strategy with its worst day absent, and the Sharpe ratio computed on it will be the best number the strategy ever produces.
The Deflated Sharpe Ratio, from Bailey and López de Prado, adds what the previous section does not: an adjustment for how many attempts produced the number in front of you.
It does not take a threshold from you. It computes the Sharpe ratio the best of N attempts would be expected to show if the true edge were zero, and asks whether yours clears that. The threshold rises with the size of the search.
Its value as a second opinion comes from the route. It works analytically on the moments of the distribution; a bootstrap test resamples the series itself. The two share few assumptions, so agreement between them is worth more than either number alone, and disagreement tells you that an assumption is carrying weight and which one to go and look at.
A strategy with negative skew produces a high proportion of winning trades by construction. Small gains are frequent and large losses are rare, so the count of winners flatters the strategy for exactly the reason the ratio does.
A win rate is therefore uninterpretable on its own. It becomes interpretable next to the average size of a win against the average size of a loss, and next to the largest single loss.
The useful trio: win rate, profit factor, worst single trade. How often, whether the wins cover the losses, and what the bad case looked like when it actually arrived rather than what a model says it should have been.
A Sharpe ratio alone is half a number. The other half is four things that fit on one line:
The length of the sample in years, because what matters is how many market regimes it covers. The skew and the kurtosis as numbers, so the reader can do the tail arithmetic above for themselves. And whether the ratio was corrected for the number of attempts behind it.
All four are cheap to compute. Which makes their absence a choice rather than an oversight.