I got handed two return streams last quarter and asked which one was riskier. Same starting capital, same number of months, and when I ran the numbers, the same standard deviation. Nearly to the second decimal. And yet one of them I would have been happy holding through a rough stretch, and the other I would not. That gap, where two things score identically on the standard measure but feel nothing alike, is the whole reason downside deviation exists. So let me walk through the arithmetic on a small example, because once you see it laid out side by side it stops being abstract.
The same dispersion, two different experiences
Standard deviation measures how far your monthly returns spread out around their own average. It does not care which direction the spread goes. A month that comes in far above average adds to the number exactly as much as a month that comes in equally far below it. That symmetry is the feature and the flaw at the same time.
Take a made-up portfolio with monthly returns of roughly plus 8, minus 2, plus 8, minus 2, repeating. The average sits at plus 3. Every month is 5 points away from that average, so the dispersion is a clean 5. Now take a second portfolio that runs plus 2, plus 4, plus 2, plus 4, and layers in one bad month of minus 6 somewhere in there. If you tune the wobble right, you can get that second stream to land on the same standard deviation as the first. To the measure, they are twins.
But look at what each one is actually doing. The first portfolio never loses much and its spread is mostly the upside overshooting. The second grinds along quietly and then hands you a real drawdown. Standard deviation charges both portfolios the same penalty for their variability, even though only one of them ever cost you anything you would notice.
How downside deviation changes the arithmetic
Downside deviation fixes the asymmetry by throwing away the good surprises. The recipe is the same as standard deviation with one edit. You pick a threshold, usually zero or your minimum acceptable return, and for every month that lands above the threshold you record a deviation of zero. Months below the threshold get measured the normal way, squared, and averaged in. Then you take the square root.
The steps look like this:
- Pick a threshold. Zero is common. Some people use the risk-free rate or a target monthly return.
- For each month, if the return is at or above the threshold, its contribution is zero.
- For each month below the threshold, take the shortfall, square it.
- Average those squared shortfalls across all periods, then take the square root.
One quiet detail trips people up. You average over the total number of periods, not just the count of losing months. Divide by the losing months only and you get something closer to the average size of a bad month, which is a different and less useful number. Dividing by all periods keeps the measure honest about how often the downside actually shows up.
Run that on my two portfolios and they finally separate. The first one, the one whose spread was all upside overshoot, has almost no months below zero, so its downside deviation collapses toward a small number. The second one keeps its minus 6 month in full, so its downside deviation stays meaningfully higher. Same standard deviation, very different downside deviation, and now the number matches the intuition I had staring at the streams.
Why this matters for asymmetric strategies
Symmetric dispersion misprices anything that is deliberately lopsided. A trend follower that takes lots of small controlled losses and occasionally catches a large winner will look volatile on standard deviation, and most of that volatility is the winners, the part you want. Penalizing it is backwards. On the other side, a strategy that sells options or picks up carry produces a long calm stretch of small positive months and then a sharp loss. Standard deviation understates that one during the calm stretch because there is so little dispersion to measure right up until the month that hurts.
Downside deviation is the engine under the Sortino ratio, the same way standard deviation sits under the Sharpe ratio. If you are comparing two managers or two of your own backtests and their Sharpe ratios are close, the Sortino ratio will often break the tie by rewarding the one whose volatility was mostly upside. When I am sifting through quant Labs results on Blockcircle, that is usually the first place I look after Sharpe, because a high Sharpe on a short window can hide a fat left tail that Sortino drags back into view.
A rule for when the simpler one is fine
Standard deviation is not wrong. It is just the right tool for a narrower job than people use it for. Here is the rule I follow.
- If the return distribution is roughly symmetric, use standard deviation. Broad index exposure, a balanced book, anything where the up months and down months mirror each other in size and frequency. The two measures will land close together and the extra math buys you nothing.
- If the strategy is deliberately asymmetric, reach for downside deviation. Trend following, options selling, anything with a skew you built on purpose. This is where the two measures diverge and the divergence is the information.
- If you do not know the shape yet, compute both. When downside deviation comes in much lower than standard deviation, your volatility is mostly upside and you can relax. When it comes in close to or above what the symmetric number implied, your pain is concentrated below the line and that is worth a second look.
The failure mode to avoid is quietly swapping one for the other across a comparison. If you rank one strategy by Sharpe and another by Sortino, you are comparing scores built on different penalties and the ranking is meaningless. Pick a measure per question and hold it constant across everything you are lining up. The measures are cheap to run. The mistake that costs you is trusting a single symmetric number to describe a payoff that was never symmetric to begin with.