In short: Standard forecasting methods return a smoothed average for slow moving items, predicting 0.7 units in a week where demand will be either zero or twelve. Classify before forecasting, using the average interval between demands and the variability of demand size when it occurs. Croston's method, from 1972, decomposes the problem into demand size and interval and updates both only when a demand occurs, and the later variants correct its bias. The decision this feeds is a stock level, which for intermittent items should come from a demand distribution rather than from a point forecast plus a buffer.
Run any standard forecasting method across a full catalogue and it will produce a number for every item. That includes the spare part that sold four times last year, the seasonal line that ships in two months and sleeps for ten, and the specialist item that goes out once a quarter in a large quantity.
The numbers those items receive are smoothed averages rather than forecasts in any meaningful sense. They will be wrong in every single period, predicting 0.7 units in a week where demand is either zero or twelve.
This is intermittent demand, and it needs different methods, different metrics and different decisions. It also tends to be a much larger share of the catalogue than anyone expects, which is why it is worth measuring rather than assuming.
Classify before you forecast
The standard classification uses two statistics computed from the demand history.
Average demand interval is the mean number of periods between non-zero demands. An item selling every week has an interval of 1. An item selling every fourth week has an interval of 4.
Squared coefficient of variation is the variance of the non-zero demand sizes divided by their squared mean. It measures how erratic the quantities are when demand does occur, ignoring the timing.
Syntetos and Boylan proposed cutoffs of roughly 1.32 on the interval and 0.49 on the squared coefficient of variation, and those two lines split the catalogue into four regions:
Smooth. Frequent demand, consistent sizes. Standard methods apply.
Intermittent. Infrequent demand, consistent sizes when it happens. Croston-family methods apply.
Erratic. Frequent demand, wildly varying sizes. Standard methods with attention to the variance, and the distribution is not going to be normal.
Lumpy. Infrequent and erratic. This is the hard quadrant, and being honest about it is more useful than pretending a method solves it.
Running this classification across your catalogue takes an hour and it changes the conversation immediately, because the answer is usually that a third to a half of active items are outside the smooth quadrant and have been receiving inappropriate treatment for years.
Kostenko and Hyndman revisited the boundary in 2006 in the Journal of the Operational Research Society and derived a different rule, drawn from where the Syntetos-Boylan approximation actually beats Croston rather than from descriptive convenience. Their line depends on the smoothing constant, so the correct cutoff for your catalogue depends on how you have parameterised the method. The two rules disagree on a modest band of items around the boundary, and the sensible response is to treat that band as unclassified and check it by hand rather than to argue about which line is right.
One data step decides whether any of this means anything. Trim leading zeros before the item's first recorded sale and trailing zeros after its last, before computing either statistic. An item listed in March and evaluated across a calendar year has nine months of genuine history and three months of zeros that carry no information, and left in place those zeros inflate the average demand interval and push the item into the lumpy quadrant. The symptom is a segment whose membership churns every month as newly listed items age in and reclassify themselves, and a planner population that stops trusting the segmentation because the same item keeps moving.
The three methods worth knowing
Croston's method, from 1972, was the first proper treatment. The insight is to decompose the problem: separately smooth the size of demand when it occurs and the interval between occurrences, then forecast the rate as size divided by interval. Both components update only when a demand occurs, so a long run of zeros does not drag the estimate down.
The result is a demand rate rather than a period-by-period prediction, which is the correct output for this kind of item. Which week the order lands stays out of reach, and the expected rate is what you need in order to set a stock level.
The Syntetos-Boylan approximation corrects a flaw in the original. Croston's estimator is biased upward, because the expectation of a ratio is not the ratio of expectations. The correction multiplies the estimate by one minus half the smoothing parameter, which is a small change with a consistent effect. Published comparisons find it beats plain Croston reliably enough that there is little reason to use the uncorrected version.
Teunter, Syntetos and Babai, usually called TSB, changes what gets updated. Instead of smoothing the interval, it smooths the probability that demand occurs, and it updates that probability every period rather than only on demand periods.
That difference matters for one specific and common situation. When an item stops selling, Croston and its variants keep their last estimate frozen, because nothing updates without a demand. The forecast stays high forever on an item that is dead. TSB's occurrence probability decays with every zero period, so the forecast declines toward zero as the evidence accumulates.
For any catalogue with genuine obsolescence, which is most catalogues, TSB is the method that stops you carrying stock recommendations for items nobody has ordered since last year.
The metric problem
This is where intermittent demand work most often goes wrong, and it is worth being firm about.
Percentage error metrics do not work here. Mean absolute percentage error requires dividing by the actual, and the actual is frequently zero, which is undefined. The usual workarounds all distort: dropping the zero periods evaluates the forecast only on the periods where something happened, which is precisely the subset where a smoothed rate looks worst.
Weighted absolute percentage error handles the zeros because it divides by total demand rather than by each observation. It is a reasonable choice for comparing across items with different volumes.
Scaled errors are better still. Mean absolute scaled error divides your error by the error of a naive baseline on the same series, so the result is a ratio where below one means you beat the baseline. Hyndman and Koehler proposed it in the International Journal of Forecasting in 2006 for exactly this situation, where percentage errors are either undefined or unstable. It is well defined on intermittent series and directly comparable across items with wildly different demand levels, which is what you need on a mixed catalogue.
There is one more thing to check that gets skipped constantly. On intermittent items, a forecast of zero everywhere will score extremely well on most error metrics, because zero is the right answer most of the time. If your reporting shows the intermittent segment with better accuracy than the fast movers, that is very likely what is happening. The comparison against a naive baseline catches it, since the all-zero forecast will not beat the baseline on the periods that matter.
Which decision this actually feeds
The real reason to get this right is the stock level, and for intermittent items the stock level should be computed from a demand distribution rather than from a point forecast plus a buffer.
Fit a count distribution. Poisson when the variance is roughly equal to the mean, negative binomial when it is larger, which is the more common case. Then compute the demand over lead time distribution by convolving the per-period distribution the required number of times, and read the order-up-to level directly off the quantile that corresponds to your economics.
Worth seeing what that changes on a specific item, because the difference is small at moderate service targets and grows quickly at high ones.
Take an item averaging 3 units a week with a weekly variance of 9, so the variance runs three times the mean and a Poisson would understate it. Lead time is 4 weeks, so demand over lead time has a mean of 12 and a variance of 36, giving a standard deviation of 6.
The normal approach reads a quantile off a z table. At 95% that is 12 plus 1.645 times 6, or 21.9, rounded to 22.
Match a negative binomial to the same mean and variance and the success probability is 12 over 36, or one third, with a shape parameter of 12 squared over 24, which is 6. Its 95th percentile is 23. One unit apart, which is no reason to change anything.
Move the target up and the two separate. At 98% the normal says 25 and the negative binomial says 27. At 99% the normal says 26 and the negative binomial says 29. That gap is the skew, and the skew sits entirely on the side you are buying protection against.
The normal has a second problem on this item that no rounding fixes. A normal with a mean of 12 and a standard deviation of 6 puts 2.3% of its probability below zero. On a 95% target that is nearly half your five point risk budget assigned to an outcome that cannot occur, and it comes out of the upper tail where the stockouts live.
That path never requires a point forecast at all, which is the correct treatment. The smoothed rate from a Croston-family method feeds the distribution parameters, and the distribution feeds the decision.
For the lumpy quadrant, even this can be more precision than the data supports, and there is a reasonable argument for reverting to simple policies. A fixed reorder point reviewed quarterly, with the review triggered by a change in demand rate rather than by a schedule, costs almost nothing to run and is not obviously worse than an elaborate model fitted to nine observations.
What I would tell a team starting this
Segment first and report the segments separately from now on. A single catalogue-wide accuracy number that mixes smooth and lumpy items is uninformative in both directions, and separating them usually reveals that the fast movers are doing better than anyone thought and the tail is doing worse.
Use TSB as the default for the intermittent segment rather than plain Croston, on the grounds that obsolescence handling is worth more than any accuracy difference between them.
Move the intermittent segment onto count distributions and stock policies, and stop reporting forecast accuracy on it as a performance measure. Report service level and stock turns instead, because those are the outcomes anyone actually cares about on these items.
The limit worth stating
For genuinely lumpy demand, the achievable accuracy is very low and no method changes that. The demand is being generated by a small number of discrete customer decisions rather than by a stable process, and the information required to predict it is in a customer's project schedule rather than in your sales history.
Where those customers are identifiable and few, the productive move is entirely outside forecasting: ask them. A quarterly conversation with the six accounts that generate the lumpy orders on a line will beat any statistical treatment, and the fact that this is not a modelling answer does not make it a worse one.
Where the demand is genuinely unpredictable and the items are cheap, the right answer is usually to hold more and stop analysing. The analysis cost exceeds the inventory cost on a large number of tail items, and a policy of generous stocking on cheap slow movers frees the planning team to work on the items where the money is.