A spare part that sells twice a year can create more planning risk than a fast-moving SKU sold every day. If planners treat both items the same, the result is predictable: excess inventory for some intermittent items, stockouts for others, and reorder points that no longer reflect reality. Learning how to forecast intermittent demand starts with recognizing that zero-demand periods are meaningful data, not noise to average away. For distributors, manufacturers, and spare-parts businesses, intermittent demand is common across long-tail assortments. It appears in replacement components, seasonal products, specialized B2B items, project materials, and low-volume e-commerce SKUs. The goal is not to force a smooth monthly forecast. It is to set replenishment parameters that protect the required service level with the least practical inventory investment. Why intermittent demand defeats standard forecasts Traditional forecasting methods work best when demand occurs regularly and volumes vary around a visible pattern. Intermittent items behave differently. A SKU may have no orders for four months, then receive one order for 12 units. The next order may arrive 45 days later for two units. A simple moving average spreads that demand across every period. It can suggest that the item sells steadily at a fraction of a unit per week, even though customers never buy it that way. When that distorted average feeds an ERP reorder point, purchasing decisions become detached from actual order behavior. The problem has two dimensions. First, demand occurrence is irregular: the question is whether an order will arrive during the planning horizon. Second, demand quantity is irregular: when an order does arrive, the size can vary sharply. A useful intermittent-demand model must address both. Lead time makes the gap more costly. If a supplier needs 10 weeks to deliver, a part with only a few historical orders may still require meaningful protection stock. Conversely, holding a large fixed buffer for every low-demand item ties up working capital without improving availability in proportion to the cost. Start with data that reflects the order pattern Forecast quality depends less on choosing a fashionable formula than on using clean transaction history and the right time structure. Start with sales-order or consumption data at the item-location level. Combining warehouses can hide local demand patterns and lead to stock being held in the wrong place. Remove or flag transactions that do not represent recurring customer demand. These may include returns, internal transfers, one-time project orders, data corrections, and exceptional bulk purchases. Do not automatically delete large orders, however. If a customer legitimately buys 100 units once a year, that event may be central to the inventory risk you need to cover. The historical window should be long enough to capture the item’s order frequency. For a part ordered three or four times per year, 24 to 36 months often provides a better view than the last six months. Older history may need less weight if the product, customer base, or market has changed. There is no single lookback period that fits every SKU. Also check whether demand is truly intermittent or simply new. A recently introduced item with only two sales has limited statistical history. It may need an analogue item, planner review, an initial service policy, or a make-to-order decision rather than a purely historical forecast. Classify items before choosing the planning method Intermittent demand should not be handled as one broad category. Classification determines how much inventory attention a SKU deserves and which replenishment logic is appropriate. ABC analysis identifies the value or revenue significance of an item. XYZ-style variability classification, demand frequency, margin, criticality, and supply lead time add the operational context. An inexpensive C item may be easy to stock generously. A low-volume A item with a long lead time may warrant a high service target because one stockout can delay a customer’s production line. This is where a blanket inventory policy fails. Setting every item to the same safety-stock rule can overprotect thousands of low-impact SKUs while underprotecting critical parts. Item-level service targets create a clearer trade-off between availability and capital. For example, a distributor might choose a 98% service target for critical repair components, 95% for standard replenishment items, and a lower target or purchase-to-order policy for slow-moving, noncritical products. Those decisions are commercial choices, not just mathematical outputs. How to forecast intermittent demand with the right model A practical intermittent-demand forecast separates the rate of demand from the size of demand. Rather than estimating a sales quantity for every period, it estimates how often orders occur and what quantity is likely when they do. Croston’s method is a well-known approach. It smooths nonzero demand sizes separately from the intervals between demand events. Variants such as SBA and TSB adjust for known bias and changing demand probability. These methods can be useful when demand is intermittent but still expected to continue. They are not always the best answer. For highly erratic or lumpy demand, the forecast may be less useful as a point estimate than a simulation of likely order events and quantities during supplier lead time. Simulation can use the actual distribution of order intervals and order sizes, rather than assuming demand follows a normal bell curve. That distinction matters. If historical orders are usually one or two units but occasionally 20, an average of four units does not describe the inventory risk very well. The reorder calculation needs to consider the chance of a large order arriving while supply is unavailable. A capable planning system can test many potential demand outcomes across the lead time and review period. It then determines the safety stock and reorder point required to meet the assigned service level. This is more realistic than applying a static percentage buffer to average demand. Convert the forecast into replenishment settings The forecast is only useful when it changes a purchasing decision. For each item-location, translate expected demand behavior into an order point, order-up-to level, minimum and maximum stock level, or another replenishment policy supported by the ERP. The reorder point should cover expected demand during lead time plus the stock needed to absorb uncertainty at the chosen service level. For intermittent demand, uncertainty is driven by both the probability of an order and the possible quantity of that order. It can also be affected by supplier reliability, review frequency, minimum order quantities, and pack sizes. Order quantity requires a separate decision. A calculated replenishment need may be three units, but a supplier may require a carton of 10 or a minimum order value across several SKUs. This is why inventory optimization cannot stop at individual item forecasts. Purchase-order optimization should consolidate supplier requirements while preventing an attractive freight break from creating years of excess stock. Consider a replacement part with a 70-day supplier lead time. It has seven orders in the past two years, typically between one and five units, with one legitimate order for 15. A planner may decide that the part needs a 97% service target because it supports a contractual repair program. A simple average may suggest only a small buffer. A lead-time simulation may show that holding six units is necessary to achieve the target, while holding 15 units adds cost with little additional service benefit. Measure forecast performance differently Mean absolute percentage error is often misleading for intermittent items. A period with zero actual demand makes percentage calculations unstable, while a forecast of one unit can appear badly wrong even when it produces the right replenishment decision. Track operational outcomes alongside forecast diagnostics. Useful measures include service level, fill rate, stockout frequency, backorder duration, inventory value, inventory turns, excess stock, and the number of emergency purchase orders. Review forecast bias over a meaningful group of items rather than reacting to every individual low-volume SKU. It also helps to monitor exceptions. An intermittent item that has not sold for far longer than its historical pattern may be becoming obsolete. An item receiving orders more frequently than expected may need a revised classification, higher service target, or supplier discussion. Planning should focus people on these changes rather than asking them to manually review every SKU every week. Keep the model connected to the ERP Intermittent-demand planning loses value when analysts export data to spreadsheets, calculate new settings, and wait weeks for changes to reach the ERP. Order history, open sales orders, purchase orders, inventory balances, lead times, and supplier constraints should be synchronized frequently. The resulting safety stock, reorder points, and purchasing recommendations must return to the operational system where buyers and planners act. ABCstock applies this workflow by classifying items, forecasting demand nightly, simulating replenishment against actual order-frequency and order-size distributions, and sending optimized parameters back to connected systems. The practical objective is straightforward: reduce safety stock where the risk does not justify it, while protecting availability where it does. The most useful forecast for an intermittent SKU is not the one that produces the neatest chart. It is the one that helps a planner answer a harder question with confidence: given this supplier lead time and this service promise, how much stock should we hold before the next unpredictable order arrives?