Crypto downside deviation measures the root mean square of return shortfalls below a declared target. An all-period calculation divides squared shortfalls by every valid observation; a downside-only calculation divides by observations below target. These measure different things. Match the target to the return interval, state the divisor and report the breach count before comparing crypto portfolios.
A positive return can still miss a positive target. A quiet period can contribute zero or a shortfall, depending on that target. Those choices explain why two reports can disagree even when they start with identical account returns.
The existing Sharpe and Sortino overview explains how downside deviation enters a performance ratio. Here the question is narrower: exactly what does the denominator measure, and how can you reproduce it?
Define the shortfall before choosing the divisor
Let r[t] be a net simple return for one equal-length period and T a fixed target in the same units. Define the nonnegative shortfall as:
- Each shortfall is
s[t] = max(T − r[t], 0). - Q is the sum of all squared shortfalls:
Q = sum(s[t]^2). - N counts all valid periods.
- M counts periods with
r[t] < T.
A return equal to target contributes zero. A return above target also contributes zero, regardless of how large the gain is. Below-target returns contribute their distance from the target, squared. Squaring gives larger misses more influence.
Both measures are root mean square (RMS) shortfalls, but they average over different sets of periods:
All-period downside deviation is sqrt(Q / N). Conditional shortfall RMS is sqrt(Q / M) when M > 0.
The official PerformanceAnalytics documentation exposes these as full and subset methods, with full the default. It also requires the minimum acceptable return, or MAR, to have the same periodicity as the returns. Record the selected method rather than relying on the label “downside deviation.”
Both calculations measure distances from the target. Neither recenters the below-target returns around their own mean. For example, two hypothetical returns of −2% and −2% have zero standard deviation between them, but their shortfall RMS against zero is 2 percentage points. Using the standard deviation of losing returns would erase their shared loss magnitude.
An eight-day calculation exposes the difference
Use this original hypothetical account-return example. Each row is a consecutive daily net simple return, with all modeled costs included and no external flows. The large daily targets are teaching choices, not suggested return objectives or observed KLINEO results.
At a 1% daily target, the worksheet is:
| Day | Net return | Shortfall below 1% target, percentage points | Squared shortfall, percentage points squared |
|---|---|---|---|
| 1 | −4% | 5 | 25 |
| 2 | −1% | 2 | 4 |
| 3 | 0% | 1 | 1 |
| 4 | +0.5% | 0.5 | 0.25 |
| 5 | +1% | 0 | 0 |
| 6 | +1.5% | 0 | 0 |
| 7 | +2% | 0 | 0 |
| 8 | +4% | 0 | 0 |
The squared shortfalls total 25 + 4 + 1 + 0.25 = 30.25. There are eight valid periods and four below-target periods. The all-period result is sqrt(30.25 / 8) = 1.9445 percentage points per daily observation, rounded. The conditional result is sqrt(30.25 / 4) = 2.7500 percentage points.
The +0.5% day belongs in the downside set because it missed 1%. The +1% day does not belong because it met the target exactly. The PerformanceAnalytics source implementation makes the strict below-target comparison and alternative counts visible; inspect the implementation and version used by your own report.
Do not mix decimal and percentage-point inputs. With returns stored as decimals, the squared sum is 0.003025, and the all-period result is 0.0194454. Displaying that decimal as a percentage gives the same 1.9445% magnitude. Squaring values expressed in percent and then treating the result as decimal returns introduces a scale error.
Frequency and severity fit an exact identity
For M > 0, let the observed breach frequency be p = M / N. Then:
All-period downside deviation equals sqrt(p) multiplied by conditional shortfall RMS.
This follows directly from Q / N = (M / N) × (Q / M). In the example, sqrt(0.5) × 2.75 = 1.9445, rounded. The all-period measure incorporates how often misses occurred and how large they were. The conditional measure describes the RMS size of misses when a miss occurred.
Suppose the hypothetical record gains eight additional valid days, each returning exactly the 1% target. No new shortfall is added. Now N = 16, M = 4 and Q = 30.25. All-period downside deviation falls to sqrt(30.25 / 16) = 1.3750; conditional shortfall RMS stays at 2.7500.
That change has an explanation: the observed frequency halved. It does not establish a reduction in the size of a bad day. Removing genuine inactive days from a calendar series can produce the reverse distortion. Missing valuations, however, are unknown observations; they must not be invented as at-target days to improve the number.
Publish the triple N, M, conditional RMS beside the all-period result. It makes frequency and severity inspectable without asking a reader to infer them from one scalar.
Raising the target can lower the conditional result
Keep all eight returns and repeat the calculation at three daily targets:
| Target | Below-target count M | Squared shortfall total Q | All-period deviation | Conditional shortfall RMS |
|---|---|---|---|---|
| 0% | 2 | 17.00 | 1.4577 | 2.9155 |
| 0.5% | 3 | 22.75 | 1.6863 | 2.7538 |
| 1% | 4 | 30.25 | 1.9445 | 2.7500 |
Shortfalls and results use percentage-point units; Q uses their square. For a fixed return sample, the all-period result cannot decrease when a fixed target rises: each squared shortfall can only increase or stay unchanged, while N stays fixed.
The conditional result can decrease because its denominator changes too. Newly included small misses dilute the conditional average. Here the zero-return day joins at the 0.5% target, and the +0.5% day joins at 1%. The lower conditional reading therefore does not mean the higher hurdle became easier to meet.
This is a useful diagnostic when comparing vendor outputs. Recompute both counts and squared totals before interpreting a counterintuitive target-sensitivity chart. The separate Omega threshold example examines unsquared gains and shortfalls; its values answer a different question.
Convert an annual target before applying it daily
A declared 12% effective annual target can be translated into a constant daily simple target using a chosen 365-day compounding convention:
The formula is T_daily = (1 + 0.12)^(1 / 365) − 1. It gives a decimal daily target of 0.0003105378, or about 0.031054% per day.
Dividing 12% by 365 instead gives about 0.032877% per day. That arithmetic allocation is a different convention. Copying 12% directly into a daily column creates a 12% daily hurdle. State whether the annual input is effective or nominal, the day count and the conversion.
Converting the target does not convert the measured downside deviation into an annual statistic. Shortfalls are thresholded before squaring, and compounding daily returns changes which annual outcomes miss an annual target. A square-root calendar multiplier is not an identity for that transformation. CFA Institute's discussion of Sortino calculation caveats highlights both divisor errors and differences between monthly and annual downside measurements.
If the question concerns annual shortfall, evaluate annual aggregated returns against the annual target when there is enough relevant history, and disclose the smaller observation count. A fitted distribution or resampling model introduces additional assumptions; it cannot turn a short observed record into evidence that all possible losses were seen.
Handle edge cases before the report reaches a reader
With valid observations but no misses, the all-period empirical result is zero. The conditional definition has no below-target observations to average, so label it unavailable under that convention. If every return equals target, the zero result still describes only that sample. With no valid observations at all, neither calculation has an estimable sample value.
When every observation misses, M = N and the two results coincide. Adding N − 1 or M − 1 silently creates another estimator; those corrections are not in the formulas used here. Record any alternative explicitly.
Even a correctly computed value loses the return sequence. Rearranging the same observations leaves downside deviation unchanged while potentially changing the drawdown path. Use a separate recovery-factor calculation when evaluating net profit relative to the largest currency drawdown, with its own capital-flow checks.
Downside deviation also supplies no win probability, payoff distribution or safe allocation. A fractional Kelly sizing analysis requires an explicit payoff model and uncertainty checks. Feeding a single downside number into a sizing rule cannot supply them.
Reproduce one result before comparing portfolios
Freeze the account-return series, valuation currency, daily boundary, costs and cash-flow treatment. Write the target and interval beside it. Then retain each shortfall, Q, N, M and the selected divisor. Recalculate the identity above and check one exactly-at-target observation.
Finally, compare portfolios only over the same window, target and return construction. Read the breach frequency and conditional severity separately, and inspect the equity path. A useful next action is to request those underlying columns from any report that provides only a downside-deviation headline.