Demand Forecasting Wiki

Syntetos–Boylan Classification (ADI and CV²): Smooth, Erratic, Intermittent, Lumpy

By MLAIA Data Science Ltd. · Published 10 November 2026 · Updated 10 November 2026

The Syntetos–Boylan classification sorts a demand series into one of four patterns using two numbers: the average demand interval, ADI, and the squared coefficient of variation of the non-zero demand sizes, CV². Cut at ADI = 1.32 and CV² = 0.49: below both cutoffs is smooth, high CV² alone is erratic, high ADI alone is intermittent, and both high is lumpy. For the 10-period series 0, 5, 0, 0, 8, 0, 12, 0, 0, 6 the ADI is 10 ÷ 4 = 2.5 and CV² is 0.12, so the series is intermittent, and the recommended estimator is SBA rather than exponential smoothing.

The two statistics

ADI — average demand interval
The mean number of periods between successive periods with non-zero demand. The simplest computation, and the one used throughout this wiki, is the number of periods in the history divided by the number of periods with demand. ADI = 1 means demand every period; ADI = 4 means demand, on average, once every four periods. (Averaging the measured gaps between demands instead gives a slightly different value on short histories, because the periods before the first demand are treated differently.)
CV² — squared coefficient of variation of demand sizes
Take only the periods with demand, compute the mean μ and the variance σ² of those sizes, and form CV² = σ² ÷ μ². It measures how variable order sizes are relative to their typical size, ignoring the zeros entirely. This wiki and the classifier tool use the population variance (divide by the number of non-zero periods); using the sample variance raises CV² slightly on short histories.

The four quadrants and the cutoffs

Demand-pattern quadrants after Syntetos, Boylan and Croston (2005), with the default forecasting method for each.
CV² < 0.49CV² ≥ 0.49
ADI < 1.32Smooth — exponential smoothing, ETS, ARIMAErratic — robust smoothing; SBA also applies
ADI ≥ 1.32Intermittent — SBA (TSB if demand may be ending)Lumpy — SBA plus an empirical lead-time-demand distribution

Where 1.32 and 0.49 come from. Syntetos, Boylan and Croston (2005) compared the theoretical mean squared error of Croston's method with that of its bias-corrected form, SBA, as functions of the demand interval and the size variability. Croston wins only when demand is frequent and stable; the boundaries of that region, approximated by two straight cuts, are ADI = 1.32 and CV² = 0.49. The scheme was therefore designed to choose between two estimators, and it has since become the standard vocabulary for describing demand patterns in general. Kostenko and Hyndman (2006) showed that the exact boundary is a single sloping line, approximately CV² > 2 − 1.5·ADI, so that for any ADI above about 1.32 SBA is preferred regardless of CV²; the two-cutoff scheme is the practical simplification of that line.

Worked example: one series per quadrant

Four short series, all computed the same way. The arithmetic for the intermittent one is shown in full; the others follow the same steps.

Intermittent — 0, 5, 0, 0, 8, 0, 12, 0, 0, 6. Ten periods, four with demand, so ADI = 10 ÷ 4 = 2.5. The non-zero sizes are 5, 8, 12 and 6, with mean 31 ÷ 4 = 7.75. Squared deviations: 7.5625, 0.0625, 18.0625 and 3.0625, summing to 28.75; variance 28.75 ÷ 4 = 7.19.

CV² = 7.19 ÷ 7.75² = 7.19 ÷ 60.06 = 0.12    → ADI 2.5 ≥ 1.32, CV² 0.12 < 0.49: intermittent

ADI and CV² for four example series, one in each quadrant. Variance is the population variance of the non-zero sizes.
SeriesPeriods / with demandADIMean of non-zero sizesVarianceCV²Quadrant
10, 12, 9, 11, 10, 136 / 61.0010.831.810.015Smooth
10, 2, 15, 30, 5, 126 / 61.0012.3380.890.53Erratic
0, 5, 0, 0, 8, 0, 12, 0, 0, 610 / 42.507.757.190.12Intermittent
0, 0, 20, 0, 1, 0, 0, 50, 0, 0, 4, 012 / 43.0018.75377.691.07Lumpy

The erratic series has demand every period, like the smooth one, but its sizes range from 2 to 30, so CV² crosses 0.49. The lumpy series shares the intermittent one's sporadic timing but its sizes range from 1 to 50, pushing CV² above 1. Paste any of these into the free classifier to reproduce the numbers.

What the quadrant tells you

Using the scheme in practice

Frequently asked questions

What do ADI and CV² stand for?

ADI is the average demand interval: the number of periods in the history divided by the number of periods that had any demand. CV² is the squared coefficient of variation of the non-zero demand sizes: their variance divided by the square of their mean. ADI measures how sporadic the timing is; CV² measures how variable the order sizes are when demand does occur.

Why are the cutoffs 1.32 and 0.49?

Syntetos, Boylan and Croston (2005) compared the expected squared error of Croston's method with that of the bias-corrected SBA estimator across demand patterns. Croston is better only when demand is both frequent and stable; the edges of that region, approximated by two straight cuts, fall at ADI = 1.32 and CV² = 0.49. Kostenko and Hyndman (2006) later showed the exact boundary is a single sloping line that the two cutoffs approximate.

How do I calculate ADI and CV² in Excel?

With the demand history in a column, ADI = COUNT(range) / COUNTIF(range, ">0"). For CV², filter or use array formulas on the non-zero values: CV² = VAR.P(non-zero values) / AVERAGE(non-zero values)^2. For the series 0, 5, 0, 0, 8, 0, 12, 0, 0, 6 this gives ADI = 10 / 4 = 2.5 and CV² = 7.19 / 7.75² = 0.12. The free demand classifier on this site does the same calculation on a pasted series.

References

See also

Try it on your own items