Demand patterns & classification

A common forecasting mistake is to run one method across every item, then blame the method when spare parts and slow movers come out badly. Experienced planners classify the demand pattern first, because the pattern decides which methods can work at all.

Two numbers do the job:

• ADI (average demand interval): total periods ÷ periods with non-zero demand. It measures how often demand happens. An item that sells every month has ADI = 1; one that sells in 4 months out of 12 has ADI = 3.
• CV² (squared coefficient of variation) of the non-zero quantities: it measures how much the size jumps around when demand does happen. Zero periods are left out on purpose; they belong to ADI.

Using the cut-offs from Syntetos, Boylan and Croston (ADI = 1.32, CV² = 0.49), every item lands in one of four boxes:

A spare part that sells 5, 80, 12 and 150 units in four random months out of twelve has ADI = 3.0 and CV² ≈ 0.9: lumpy. No point forecast will predict it well, so the sensible move is to manage it through stock policy and a forecast of the range of demand.

Computing the two numbers

For each item, take a fixed stretch of history, say 12 months:

1. ADI = number of periods ÷ number of periods with demand.
2. Take only the non-zero quantities, compute their mean $\mu$ and standard deviation $\sigma$, then $CV^2 = (\sigma/\mu)^2$.
3. Place the item using the cut-offs ADI = 1.32 and $CV^2$ = 0.49.

Here $\sigma$ is the plain (population) standard deviation of the non-zero quantities. Some tools use the sample version, which gives slightly higher values, so stick to one convention.

From pattern to method

For smooth items, check trend and seasonality next. Plot the series and compare each month’s average with the overall average: a seasonal index of 1.3 means a month runs 30% above normal.

The time bucket changes the answer

The same item can be lumpy week by week and smooth month by month. Adding periods together (temporal aggregation) reduces the zeros and smooths the sizes. Some methods exploit this on purpose: forecast at a coarser bucket where the pattern is easier, then split the result back down (the ADIDA approach). Choose the bucket that matches the decision: weekly replenishment needs weekly patterns, while capacity planning can use monthly ones.

Look behind the pattern

Erratic or lumpy demand often has a visible cause:

• A few large customers: one customer ordering 150 units twice a year makes an item lumpy even if many small customers buy steadily. Forecasting that customer separately, or simply asking about their plans, turns noise into information.
• Order batching: customers who order monthly instead of weekly create lumps out of smooth consumption. Their sell-through or usage data removes the lumps.
• Promotions and events: spikes that should be modelled as uplift, not left in as variability.

Before choosing a method for an awkward item, ask why it is awkward.

Patterns change

Items move between classes as they grow, mature and decline: a new product may be intermittent at launch, smooth at maturity and intermittent again before phase-out. A structural break, such as a lost contract or a new channel, can change the pattern overnight. Re-classify regularly, and use only recent history after a known break.

Where the cut-offs come from

The values 1.32 and 0.49 were derived by comparing the theoretical errors of Croston’s method, its bias-corrected version and exponential smoothing: they mark the regions where each tends to do best. They are sensible defaults, not physical constants, and many teams adjust them after checking which method actually wins on their own data.