The problem with forecasting zeros
Most forecasting methods assume demand happens every period. Exponential smoothing tracks a level that updates with each new observation. ARIMA models fit autoregressive structure to a continuous signal. Neural networks learn patterns from dense historical data. All of them struggle when 60% of your periods have zero demand, because there is nothing to smooth, nothing to correlate, and nothing to learn from in those gaps.
This is the reality for a large fraction of manufacturing SKUs. Spare parts, service items, slow-moving components, long-tail finished goods, and new products before they ramp all exhibit intermittent demand. For some manufacturers, intermittent SKUs are 40% or more of the item master. Forecasting them badly means carrying too much inventory on items that rarely sell, or stocking out when they finally do.
In 1972, J.D. Croston published a method specifically designed for this problem. It is still the most widely taught and most widely implemented approach. But it is not the last word. This article covers what Croston's method actually does, where it fails, and what the research since 1972 has produced that you should consider instead.
What Croston's method actually does
Croston's insight was simple and correct. He separated the intermittent demand problem into two sub-problems. First, when will the next demand occur? Second, when it does occur, how big will it be?
The method maintains two separate exponential smoothing models. One tracks the time between demand occurrences (the inter-demand interval). The other tracks the size of demand when it does occur. The forecast for any given period is the expected demand size divided by the expected inter-demand interval. If demand occurs on average every 4 periods and averages 100 units when it does, the per-period forecast is 25 units.
This separation matters because it respects the structure of intermittent demand. A naive exponential smoothing model applied to intermittent data will produce a forecast that decays toward zero during gaps and spikes when demand occurs. Croston's method produces a steady, low per-period forecast that better reflects the true expected demand rate. The original paper, published in the Operational Research Quarterly, demonstrated this empirically on inventory data.
Where Croston fails
Croston's method is not wrong. It is incomplete. Three specific issues have been documented in the subsequent literature.
First, Croston's method is biased. The original method uses simple exponential smoothing on the inter-demand interval, which introduces a positive bias in the forecast. The method over-forecasts demand rate, particularly for items with long inter-demand intervals. This was identified formally by Syntetos and Boylan in a 2001 paper published in the International Journal of Forecasting, and the bias correction they proposed became the SBA (Syntetos-Boyland Approximation) method, which is now the most widely recommended improvement over original Croston.
Second, Croston assumes demand sizes and inter-demand intervals are independent. In practice, they often are not. Large demand occurrences may be followed by longer gaps (a customer bought a bulk order and will not buy again for a while). Or the opposite. The independence assumption is a simplification that works acceptably for many SKUs but fails for others. More recent methods, including those based on aggregate hierarchical structure, attempt to model this dependence.
Third, Croston does not handle obsolescence. An item that has not had demand in 18 months is either truly dead or due for a resurgence. Croston's method will keep producing a small non-zero forecast indefinitely, which leads to inventory being held against items that will never sell again. Detection of demand extinction is a separate problem that Croston does not address.
What came after Croston
The intermittent demand forecasting literature since 1972 is substantial. Three methods are worth knowing about as a practitioner.
SBA (Syntetos-Boylan Approximation). Published in 2005 in the European Journal of Operational Research, SBA applies a bias correction to Croston's method. The correction is a single multiplicative factor applied to the forecast. SBA is simple to implement on top of an existing Croston implementation, and the empirical evidence across multiple studies shows it reduces forecast error modestly but consistently compared to original Croston. A comparison study published in the International Journal of Production Economics tested SBA against Croston and simple exponential smoothing across multiple real datasets and found SBA superior in the majority of cases. If you are using Croston today, SBA is a near-free upgrade.
TSB (Teunter-Syntetos-Babai). Published in 2011 in the European Journal of Operational Research, TSB replaces the inter-demand interval model with a direct probability model for demand occurrence. Instead of tracking time between demands, it tracks the probability that demand occurs in any given period. This handles obsolescence better, because the probability can decay toward zero during long gaps. TSB is more complex to implement than SBA but addresses the obsolescence problem that both Croston and SBA leave open.
Hierarchical and cross-sectional methods. More recent research, including work by Hyndman and colleagues, applies hierarchical forecasting to intermittent demand. The idea is to forecast at an aggregate level (product family, location group) where demand is denser, then disaggregate to the SKU level using proportional allocation. This can outperform per-SKU methods when the hierarchy is well-defined and the aggregate signal is strong. The M5 forecasting competition, results published in the International Journal of Forecasting in 2022, demonstrated that hierarchical and ensemble approaches consistently outperformed single-method approaches across a large retail dataset.
What to actually use
For most manufacturers, the practical decision tree is this.
If your intermittent SKUs are a small fraction of your item master and your demand is not heavily skewed, SBA is a safe default. It is a documented improvement over Croston, it is implemented in most modern forecasting libraries, and it requires no additional data.
If obsolescence is a real concern (long-tail spare parts, end-of-life products), TSB is worth the implementation effort. The probability-of-occurrence model handles the decay toward zero that Croston and SBA miss.
If you have a strong product hierarchy and dense aggregate demand, hierarchical forecasting can be layered on top of either SBA or TSB. The M5 results suggest this is where the biggest accuracy gains are available, but it requires more sophisticated tooling.
Neural network approaches (DeepAR, N-BEATS applied to intermittent data) are an active research area. The M5 competition showed that neural methods performed well on dense, high-volume series but did not consistently beat statistical methods on sparse, intermittent series. For intermittent demand specifically, the simpler statistical methods remain the practical choice in 2026.
The honest summary
Croston's method was a genuine advance in 1972 and it is still a reasonable starting point in 2026. But 54 years of research have produced documented improvements. If your forecasting tool ships original Croston without an SBA option, it is a generation behind the literature. If it ships SBA without an obsolescence model, it is half a generation behind.
For the practitioner, the question is not which method is theoretically best. The question is which method your tool supports, whether your demand patterns actually exhibit the bias and obsolescence problems these methods address, and whether the accuracy gain justifies the implementation and tuning effort. Start with SBA. Move to TSB if obsolescence is a real cost. Layer hierarchical forecasting on top if your data supports it. Measure the result with WMAPE, not MAPE, because intermittent demand is exactly the case where MAPE misleads.