Sortino vs. Sharpe: Two Ways to Divide by Risk

Sharpe and Sortino share a numerator and disagree about the denominator. Both divide excess return by a measure of risk; Sharpe uses the standard deviation of all returns, while Sortino uses the deviation of only the returns below a target. The argument for Sortino is that volatility on the upside isn’t risk, and penalizing it treats a good surprise as a bad one. The argument against is that the modification introduces its own assumptions and makes the estimate noisier.

Educational material, not trading advice. Algorithmic crypto trading is high-risk and most retail algo traders lose money.

Where they differ, precisely

Sharpe’s denominator is the standard deviation of the full return series; Sortino’s is the downside deviation relative to a target. The Sortino denominator, written for a target T:

downside_dev = sqrt( mean( min(r - T, 0)^2 ) )
Sortino      = mean(r - T) / downside_dev

Two implementation subtleties do most of the damage in practice.

The target is a choice, not a constant. Setting T to zero measures deviation below break-even. Setting it to the risk-free rate measures deviation below the safe alternative. Setting it to a “minimum acceptable return” measures deviation below an aspiration. All three are used, they give different numbers, and a Sortino ratio without its target stated is not reproducible.

The averaging convention differs between implementations. Some divide the sum of squared shortfalls by the count of all periods; others divide by the count of below-target periods only. The first is the standard definition and treats a rare shortfall as rare; the second is a conditional average and produces a systematically smaller denominator, hence a larger ratio. Libraries disagree. Check the source before comparing your number to anyone else’s.

The case for Sortino

It matches how risk is actually experienced. A strategy whose returns are mostly small and occasionally very large has high standard deviation entirely because of the good outcomes. Sharpe penalizes it; a reader who cares about losses would not. This is not a hypothetical shape — trend-following and other convex, long-optionality strategies characteristically look mediocre on Sharpe and better on Sortino, because their return distribution is right-skewed by design. See momentum vs. mean-reversion for why that skew arises from the strategy logic rather than from luck.

It exposes asymmetry that Sharpe averages away. Comparing the two ratios for the same strategy is more informative than either alone. If Sortino is much higher than Sharpe, the return distribution is right-skewed: the volatility is mostly upside. If they are close, the distribution is roughly symmetric. If Sortino is lower, the downside dominates — a strategy whose bad periods are worse than its good ones are good, which is exactly the profile that a headline Sharpe can conceal.

The case against

Sortino throws away half the sample, so the estimate is noisier. Downside deviation is computed from below-target observations only. With daily data over a short history that can be a small number of points, and a statistic built on few observations has a wide confidence interval. The sampling-error problem that already afflicts Sharpe is worse here, not better.

It is easier to game. A strategy can be constructed to have very few small losses and one catastrophic one — writing insurance, in effect. Until the catastrophe, downside deviation is tiny and Sortino is spectacular. Sharpe is somewhat less flattering to that shape, and neither metric catches it properly, because both are second-moment measures and the risk lives in the tail. If a ratio is your selection criterion, you are selecting for whatever shape maximizes that ratio, and a search procedure will find it — the general hazard described in how overfitting hides in a trading strategy.

It doesn’t measure path. Like Sharpe, Sortino is order-independent: shuffle the returns and the ratio is unchanged. Clustering of losses is what produces a deep drawdown, and no ratio of moments sees clustering. That job belongs to maximum drawdown.

Neither one is the risk measure

The honest position is that both are summary statistics of the second moment, and crypto’s risk is largely in the fourth. Fat tails mean the rare, extreme observation dominates the real outcome while contributing modestly to either denominator. A metric that squares deviations does register large moves more heavily than small ones, but it still assumes the distribution is describable by its spread — and for a series with occasional enormous moves, it isn’t.

What that implies practically: use these ratios for comparing candidates within a family, not for deciding whether a strategy is safe. Safety questions need tail-oriented answers — worst observed period, drawdown depth and duration, behaviour in the worst decile of days, and what happens under a deliberately punitive cost and gap assumption.

A practical reporting habit

Report both, with their parameters, and one path-aware measure alongside. A defensible minimum:

• Annualized Sharpe, with the periods-per-year constant and the risk-free assumption stated.
• Annualized Sortino, with the target and the averaging convention stated.
• Maximum drawdown and its duration, because neither ratio sees it.
• The sample length and number of below-target periods, so a reader can judge how much either estimate is worth.

And a rule that saves time: pick one of the two as your optimization objective before you start searching, and keep the other purely as a diagnostic. Optimizing whichever looks better after the fact is metric shopping, and it inflates results in exactly the way that survives a walk-forward split — the tuning is clean, but the choice of what to tune was made with hindsight.

The takeaway

Sortino is a reasonable refinement with a real motivation: upside variation is not the thing you’re trying to avoid, and a symmetric denominator pretends otherwise. It is most worth using when you have a specific reason to expect skew, and it earns its keep less as a replacement for Sharpe than as a comparison against it — the gap between the two ratios is the actual information.

What neither ratio will do is tell you whether an account survives. That depends on sizing, on the clustering of losses, and on how the tails behave when the market does something it hasn’t done in your sample. For that side of the problem, see position sizing and risk management basics — the arithmetic there outranks any number of decimal places in a ratio.