Semideviation and Downside Risk

Imagine two managers with identical Volatility. Manager A's bad months are as bad as her good months are good, a symmetric coin flip around the average. Manager B's good months are wildly good but her bad months are small and rare. Standard deviation scores them the same. Every investor on the planet would rather hold Manager B. That mismatch is the whole reason semideviation exists: ordinary volatility treats a surprise $50,000 gain as exactly as "risky" as a $50,000 loss, when only one of those keeps you up at night.

The analogy

Think of a parent judging how safe a playground is. They do not care how high the swings go if the child is having fun going up. They care about how far the child could fall. A safety score that averaged "distance travelled up" and "distance travelled down" would badly mis-rank playgrounds. The parent's real question is: if things go wrong, how far wrong, and how often? Semideviation is that question asked of a return series, ignoring every month that went right.

Building the formula, one piece at a time

Start from a target return $\tau$, often zero or the risk-free rate, the line between "fine" and "bad." For each historical return $R_t$, define the shortfall as $\min(R_t - \tau,\ 0)$: if the return beat the target, the shortfall is zero (no penalty); if it missed, the shortfall is the negative gap.

Square each shortfall, average over $n$ observations, and take the square root:

$\text{SemiDev} = \sqrt{\frac{1}{n}\sum_{t=1}^{n} \min(R_t - \tau,\, 0)^2}.$

In words: look only at the months that undershot the target, measure how far each one undershot, square those distances (so bigger misses count much more than small ones), average them, then undo the squaring with a square root so the answer is back in return units. Every good month contributes exactly zero — it is not "cancelled out" against a bad month, it is simply invisible to the formula, which is precisely the property ordinary standard deviation lacks.

Drag the mean of this bell curve up and down. Ordinary volatility measures the spread on both sides equally. Semideviation is what you would get if you covered the right half of the curve with your hand and only measured the spread of what is left.

Worked example 1: five monthly returns

A fund posts monthly returns of +4%, -2%, +1%, -5%, +3%, target $\tau = 0$.

Shortfalls: $\min(4,0)=0$, $\min(-2,0)=-2$, $\min(1,0)=0$, $\min(-5,0)=-5$, $\min(3,0)=0$.

Square them: $0,\ 4,\ 0,\ 25,\ 0$. Average: $(0+4+0+25+0)/5 = 5.8$. Square root: $\sqrt{5.8} \approx 2.41\%$.

Compare to ordinary standard deviation of the same series: the mean is $0.2\%$, squared deviations from the mean are roughly $14.4, 4.8, 0.64, 27.0, 7.8$, averaging to $10.9$, giving $\sqrt{10.9} \approx 3.30\%$. Semideviation (2.41%) is meaningfully smaller than full deviation (3.30%) here, because two of the five months (the two big positive ones) were penalized by ordinary volatility for no real reason.

Worked example 2: same volatility, different shape

Manager A: returns $-3\%, +3\%, -3\%, +3\%$ (perfectly symmetric). Standard deviation and semideviation must agree here, since every negative return has a matching positive one of equal size: semideviation is $\sqrt{(9+0+9+0)/4} = \sqrt{4.5} \approx 2.12\%$, and ordinary deviation on this zero-mean series is the same $2.12\%$.

Manager B: returns $-1\%, +7\%, -1\%, +1\%$, similar spread by eye. Semideviation around $\tau=0$: shortfalls are $-1, 0, -1, 0$, squared $1,0,1,0$, average $0.5$, root $\approx 0.71\%$. Manager B's downside risk is three times smaller than Manager A's despite "bouncing around" a similar amount, because B's bad days are small and its one big move is a gain. A Sharpe ratio comparison would miss this; a Sortino ratio catches it immediately.

What this means in practice

Desks that report both a Sharpe and a Sortino ratio are flagging exactly this gap: a strategy with a mediocre Sharpe but a strong Sortino has lopsided, favorable-shaped returns (common in trend-following and some option-selling books, at least until the tail arrives). Semideviation also feeds directly into target-shortfall risk budgeting, where a pension fund cares only about the probability and size of missing its liability, not about upside variance it is happy to keep.